Design Guidelines for Deploying Closed Loop Systems

Transcription

ANALYSIS OF FINGERPRINT SENSOR INTEROPERABILITY ON SYSTEM
PERFORMANCE
A Dissertation
Submitted to the Faculty
of
Purdue University
by
Shimon K. Modi
In Partial Fulfillment of the
Requirements for the Degree
of
Doctor of Philosophy
August 2008
Purdue University
West Lafayette, Indiana
ii
TABLE OF CONTENTS
Page
CHAPTER 1. INTRODUCTION ............................................................................ 1
1.1. Introduction ................................................................................................. 1
1.2. Statement of Problem ................................................................................. 3
1.3. Significance of Problem .............................................................................. 4
1.4. Purpose of Study ........................................................................................ 7
1.5. Definitions ................................................................................................... 8
1.5.1. Sensor Related Distortions and Variations ......................................... 12
1.6. Assumptions ............................................................................................. 14
1.7. Delimitations ............................................................................................. 14
1.8. Limitations ................................................................................................ 15
1.9. Summary .................................................................................................. 15
CHAPTER 2. REVIEW OF LITERATURE .......................................................... 16
2.1. Introduction ............................................................................................... 16
2.2. History of Biometrics ................................................................................. 16
2.2.1. Studies Related to Finger Anatomy and Ridge Structure ................... 16
2.2.2. Early Experiments with Fingerprint Identification: Criminalistics ......... 17
2.2.3. Galton’s Fingerprint Experiments ....................................................... 18
2.2.4. Henry’s Fingerprint Experiments ........................................................ 19
2.2.5. Fingerprint Recognition in South America, Europe ............................. 20
2.3. Automated Fingerprint Recognition .......................................................... 21
2.4. General Biometric Recognition System .................................................... 21
2.4.1. Fingerprint Recognition Model ............................................................ 25
2.4.2. Fingerprint Acquisition Technologies .................................................. 27
2.4.3. Fingerprint Image Quality Assessment ............................................... 34
2.4.4. Fingerprint Feature Extraction ............................................................ 39
2.4.5. Fingerprint Matching ........................................................................... 45
2.5. Large Scale Systems and Distributed Authentication Architectures ......... 47
2.5.1. Fingerprint System Interoperability Experiments ................................ 49
2.6. Performance Metrics ................................................................................ 55
2.6.1. Receiver Operating Characteristic Curves.......................................... 56
2.6.2. Equal Error Rate ................................................................................. 56
2.6.3. Difference in Match and Non-Match Scores ....................................... 57
2.6.4. User Probabilities and Cost Functions ................................................ 57
2.6.5. Cumulative Match Curve .................................................................... 58
iii
CHAPTER 3. METHODOLOGY ......................................................................... 59
3.1. Introduction ............................................................................................... 59
3.1.1. Research Design ................................................................................ 59
3.2. Data Collection Methodology .................................................................... 60
3.2.1. Participant Selection ........................................................................... 61
3.2.2. Timeline .............................................................................................. 61
3.2.3. Data Collection Hardware & Software ................................................ 61
3.2.4. Fingerprint Sensors ............................................................................ 63
3.2.5. Fingerprint Sensor Maintenance......................................................... 63
3.2.6. Software or Sensor Malfunction.......................................................... 64
3.2.7. Variables Measured during Data Collection........................................ 64
3.2.8. Variables Controlled during Data Collection ....................................... 65
3.3. Data Processing Methodology .................................................................. 65
3.3.1. Minutiae Count and Image Quality Processing ................................... 65
3.3.2. Fingerprint Feature Extractor and Matcher ......................................... 66
3.4. Data Analysis Methodology ...................................................................... 66
3.4.1. Score Generation Methodology .......................................................... 66
3.4.2. Analysis Techniques ........................................................................... 69
3.4.3. Impact of Image Quality on Interoperable Datasets............................ 76
3.4.4. Post Hoc Analysis ............................................................................... 76
3.5. Threats to Internal and External Validity ................................................... 77
3.5.1. Internal Validity ................................................................................... 77
3.5.2. External Validity .................................................................................. 80
3.6. Evaluation Classification ........................................................................... 81
3.7. Summary .................................................................................................. 82
CHAPTER 4. DATA ANALYSIS ......................................................................... 83
4.1. Failure to Enroll (FTE) .............................................................................. 84
4.2. Basic Fingerprint Feature Analysis ........................................................... 85
4.2.1. Minutiae Count Analysis ..................................................................... 85
4.2.2. Image Quality Analysis ....................................................................... 95
4.3. Match Score Analysis ............................................................................. 104
4.3.1. VeriFinger 5.0 Interoperability Error Rates ....................................... 104
4.3.2. NBIS Interoperability Error Rates ..................................................... 109
4.3.3. Test of Proportions for VeriFinger 5.0 Match Scores ........................ 111
4.3.4. Test of Proportions for BOZORTH3 Match Scores ........................... 115
4.3.5. Test of Similarity of VeriFinger 5.0 Match Scores ............................. 118
4.4. Impact of Quality on Interoperability ....................................................... 121
4.5. Acquisition and Interaction Level Interoperability.................................... 128
4.6. Investigative Analysis ............................................................................. 129
4.7. Fingerprint Image Transformation .......................................................... 133
CHAPTER 5. CONCLUSIONS AND RECOMMENDATIONS........................... 142
5.1. Conclusions ............................................................................................ 142
5.2. Contributions .......................................................................................... 144
5.3. Future Work ............................................................................................ 144
LIST OF REFERENCES .................................................................................. 147
iv
Appendix A. ................................................................................................... 153
Appendix B. ................................................................................................... 154
Appendix C. ................................................................................................... 155
Appendix D. ................................................................................................... 156
Appendix E. ................................................................................................... 157
Appendix F. ................................................................................................... 160
Appendix G. ................................................................................................... 161
Appendix H. ................................................................................................... 164
Appendix I...................................................................................................... 167
Appendix J. .................................................................................................... 169
VITA ................................................................................................................. 172
v
LIST OF TABLES
Table
Page
Table 2.1 Possible Acquisition/Storage/Matching Location ................................ 26
Table 3.1 List of Fingerprint Sensors .................................................................. 63
Table 3.2 Evaluation Classification (Mansfield & Wayman, 2002). ..................... 82
Table 4.1 Coding of Fingerprint Sensors and Datasets ...................................... 83
Table 4.2 Number of participants that recorded Failure to Enroll (FTE) using
VeriFinger 5.0 ............................................................................................... 84
Table 4.3 Descriptive statistics for Aware software minutiae count analysis ...... 85
Table 4.4 p values for Tukey’s HSD test for pairwise dataset effects for Aware
software minutiae count ................................................................................ 88
Table 4.5 Descriptive Statistics for MINDTCT minutiae count ............................ 89
Table 4.6 p values for Tukey’s HSD test for pairwise dataset effects for
MINDTCT minutiae count ............................................................................. 91
Table 4.7 Descriptive statistics for VeriFinger 5.0 minutiae count ...................... 92
Table 4.8 p values for Tukey’s HSD pairwise test for dataset effects ................. 94
Table 4.9 Descriptive statistics for Aware software quality scores...................... 95
Table 4.10 Summary of Sensor Acquisition Technologies and Interaction Types
..................................................................................................................... 98
Table 4.11 p values for test for difference in quality scores in every possible
pairwise comparison using Aware Quality scores......................................... 99
Table 4.12 Descriptive Statistics for NFIQ Image Quality Analysis................... 100
Table 4.13 p values for test for difference in quality scores in every possible
pairwise comparison using NFIQ scores .................................................... 103
Table 4.14 FNMR at FMR 0.01% (in percentage values) ................................. 105
Table 4.15 FNMR at FMR 0.1% (in percentage values) ................................... 106
Table 4.16 Percentage of fingerprint pairs with detected core .......................... 108
Table 4.17 FNMR (at threshold of score of 40)................................................. 110
Table 4.18 Results of Overall Test of Proportions ............................................ 112
Table 4.19 p values of pairwise test of proportions........................................... 114
Table 4.20 Results of overall test of proportions............................................... 116
Table 4.21 p values of pairwise test of proportions........................................... 117
Table 4.22 Results of overall test of similarity of genuine match scores........... 119
Table 4.23 Results of overall test of similarity of imposter match scores.......... 120
Table 4.24 Number of comparisons in full datasets .......................................... 122
Table 4.25 Number of comparisons for datasets with images of NFIQ < 4....... 123
vi
Table 4.27 Results of overall test of proportions............................................... 125
Table 4.28 Results of pairwise test of proportions ............................................ 127
Table 4.30 Correlation matrix of match score, quality score difference, ridge
bifurcation count difference, ridge ending count difference ........................ 129
Table 4.31 Recalculated FNMR (FMR = 0.1%) ................................................ 138
Table 4.32 FNMR for S3 dataset (FMR=0.1%)................................................. 138
Table I.1 Results of Survey .............................................................................. 168
Table J.1 Correlation matrix of Quality Score, Temperature, Moisture Content,
Oiliness, Elasticity....................................................................................... 169
vii
LIST OF FIGURES
Figure
Page
Figure 1.1 Example of an Interoperability Scenario. ............................................. 4
Figure 1.2 Annual Industry Revenue Projections 2007-2012 (IBG, 2007). ........... 6
Figure 1.3 Biometric Market Share by Technology (IBG, 2007). .......................... 7
Figure 1.4 Purpose of Study. ................................................................................ 8
Figure 2.1 Examples of Galton features. ............................................................ 19
Figure 2.2 The General Biometric System (Wayman, 1997) .............................. 23
Figure 2.3 Biometric Authentication System (R. M. Bolle, Connell, Pankanti,
Ratha, & Senior, 2004b) ............................................................................... 24
Figure 2.4 ISO/IEC JTC1 SC37 Conceptual Biometric Model (ISO)................... 25
Figure 2.5 Possible Distributed Architectures ..................................................... 26
Figure 2.6 Optical Fingerprint Sensor ................................................................. 29
Figure 2.7 Fingerprint Image from Optical Fingerprint Sensor ............................ 29
Figure 2.8 Capacitive Fingerprint Sensor ........................................................... 31
Figure 2.9 Fingerprint Image from Capacitive Fingerprint Sensor ...................... 32
Figure 2.10 Thermal Fingerprint Sensor ............................................................. 33
Figure 2.11 Fingerprint Image from Thermal Fingerprint Sensor ........................ 33
Figure 2.12 3X3 matrix of P (Boashash et. al, 1997) .......................................... 41
Figure 2.13 Pixel pattern used to detect ridge endings (Garris et al., 2004) ....... 44
Figure 2.14 Pixel patterns for minutiae detection (Garris et al., 2004) ................ 44
Figure 2.15 Unnecessary features (Garris et al., 2004) ...................................... 45
Figure 3.1 Triplesense TR-3 ............................................................................... 62
Figure 3.2 Raytek MiniTempTM ........................................................................... 62
Figure 3.3 Generation of match scores .............................................................. 68
Figure 3.4 Scores of native and interoperable datasets ..................................... 69
Figure 3.5 Performance interoperability matrix at different operational FMR ...... 73
Figure 4.1 Boxplot of Aware software minutiae count......................................... 86
Figure 4.2 Histogram of Aware software minutiae count .................................... 86
Figure 4.3 Boxplot of Minutiae Count Extracted by MINDTCT ........................... 90
Figure 4.4 Histogram of Minutiae Count Extracted by MINDTCT ....................... 90
Figure 4.5 Boxplot of Minutiae Count Extracted by VeriFinger 5.0 ..................... 93
Figure 4.6 Histogram of Minutiae Count Extracted by VeriFinger 5.0 ................. 93
Figure 4.7 Boxplot of Image Quality Scores Computed by Aware ...................... 96
Figure 4.8 Histogram of Image Quality Scores Computed by Aware.................. 96
Figure 4.9 Normality plot of residuals ................................................................. 97
Figure 4.10 Boxplot of Image Quality Scores Computed by NFIQ.................... 100
viii
Figure 4.11 Histogram of Image Quality Scores Computed by NFIQ ............... 101
Figure 4.12 Scatter plot of Consistency of Placement vs. FNMR (FMR=0.1%) 109
Figure 4.13 Normality plot of residuals of S1 genuine match scores ................ 118
Figure 4.14 Normality plot of residuals of S1 imposter match scores ............... 120
Figure 4.15 Scatter plot of VeriFinger 5.0 Match Score vs. Difference in Quality
Scores ........................................................................................................ 130
Figure 4.16 Scatter plot of VeriFinger 5.0 Match Score vs. Difference in Ridge
Bifurcation Count ........................................................................................ 131
Ending ........................................................................................................ 132
Figure 4.18 Skeletonized fingerprint image from S8(left) and S3(right) ............ 134
Figure 4.19 Traversal path for IMPROFILE ...................................................... 135
Figure 4.20 IMPROFILE output for S8 dataset image ...................................... 136
Figure 4.21 IMPROFILE output for S3 dataset image ...................................... 136
Figure 4.22 Fingerprint image after transformation........................................... 137
Figure 4.23 Ridge spacing profile along x-axis. S1 image on left and S6 image on
right ............................................................................................................ 139
Figure 4.24 Ridge spacing profile along y-axis. S1 mage on left and S6 image on
right ............................................................................................................ 140
Figure 5.1 Eight quadrant ridge spacing profile ................................................ 146
Figure A.1 Data Collection Protocol.................................................................. 153
Figure C.1 Flowchart of fingerprint feature statistical analysis .......................... 155
Figure D.1 Physical layout of data collection area ............................................ 156
Figure E.1 Normality plot of residuals of Aware Minutiae Count ....................... 157
Figure E.2 Residuals vs. Fitted values of Aware Minutiae Count ..................... 157
Figure E.3 Time series plot of observations ..................................................... 158
Figure E.4 Normality plot of residuals of NBIS Minutiae Count ........................ 158
Figure E.5 Residuals vs. Fitted values of NBIS Minutiae Count ....................... 159
Figure E.6 Time series plot of observations ..................................................... 159
Figure F.1 Normality plot of residuals of Aware Quality Scores ........................ 160
Figure G.1 Normality plot of residuals of match scores from imposter
comparisons with S1 dataset ...................................................................... 161
ix
Figure H.1 Normality plot of residuals of match scores from genuine
Figure J.1 Scatter plot of Aware Quality Score vs. Temperature ..................... 170
Figure J.2 Scatter plot of Aware Quality Score vs. Moisture ............................. 170
Figure J.3 Scatter plot of Aware Quality Score vs. Oiliness .............................. 171
Figure J.4 Scatter plot of Aware Quality Score vs. Elasticity ............................ 171
x
ABBREVIATIONS
AFIS - Automated Fingerprint Identification System
ANOVA - Analysis of Variance
FMR - False Match Rate
FNMR – False Non Match Rate
FTE – Failure to Enroll
H0 – Null Hypothesis
H1 – Alternate Hypothesis
ISO – International Organization for Standardization
M – Mathematical mean
NBIS – NIST Biometric Image Software
NFIQ – NIST Fingerprint Image Quality
NIST- National Institute of Standards and Technology
ROC – Receiver Operating Characteristic
SD – Mathematical standard deviation
WSQ – Wavelet Scalar Quantization
xi
ABSTRACT
Modi, Shimon K. Ph.D., Purdue University, August, 2008. Analysis of fingerprint
sensor interoperability on system performance. Major Professor: Stephen Elliott.
The increased use of fingerprint recognition systems has brought the issue of
fingerprint sensor interoperability to the forefront. Fingerprint sensor
interoperability refers to the process of matching fingerprints collected from
different sensors. Variability in the fingerprint image is introduced due to the
differences in acquisition technology and interaction with the sensor. The effect
of sensor interoperability on performance of minutiae based matchers is
examined in this dissertation. Fingerprints from 190 participants were collected
on nine different fingerprint sensors which included optical, capacitive, and
thermal acquisition technologies and touch, and swipe interaction types. The
NBIS and VeriFinger 5.0 feature extractor and matcher were used. Along with
fingerprints, characteristics like moisture content, oiliness, elasticity and
temperature of the skin were also measured. A statistical analysis framework for
testing interoperability was formulated for this dissertation, which included
parametric and non-parametric tests. The statistical analysis framework tested
similarity of minutiae count, image quality and similarity of performance between
native and interoperable datasets. False non-match rate (FNMR) was used as
the performance metric in this dissertation. Interoperability performance analysis
was conducted on each sensor dataset and also by grouping datasets based on
the acquisition technology and interaction type of the acquisition sensor.
Similarity of minutiae count and image quality scores between two datasets was
not an indicator of similarity of FNMR for their interoperable datasets.
xii
Interoperable FNMR of 1.47% at fixed FMR of 0.1% was observed for the optical
touch and capacitive touch groupings. The impact of removing low quality
fingerprint images on the effect of interoperable FNMR was also examined.
Although the absolute value of FNMR reduced for all the datasets, fewer
interoperable datasets were found to be statistically similar to the native datasets.
An image transformation method was also proposed to compensate for the
differences in the fingerprint images between two datasets, and experiments
conducted using this method showed significant reduction in interoperable FNMR
using the transformed dataset.
1
CHAPTER 1. INTRODUCTION
1.1. Introduction
Authentication of individuals is a process that has been performed in one form or
another since the beginning of recorded history. Evidence of attempts to
authenticate individuals has been found dating back to 500 B.C (Ashbourn,
2000). The quest for improving authentication methodologies has intrigued
humans since these early times. Establishing and maintaining the identity of
individuals, and accurate automated recognition is becoming increasingly
important in today’s networked world. As technology advances, the complexity of
these tasks has also increased. There are three main methods of authenticating
an individual: 1) using something that only the authorized individual has
knowledge of e.g. passwords 2) using something that only the authorized
individual has possession of e.g. physical tokens 3) using physiological or
behavioral characteristics that only the authorized individual can reproduce i.e.
biometrics. The increasing use of information technology systems has created
the concept of digital identities which can be used in any of these authentication
mechanisms. Digital identities and electronic credentialing have changed the way
authentication architectures are designed. Instead of stand-alone and monolithic
authentication architectures of the past, today’s networked world offers the
advantage of distributed and federated authentication architectures. The
development of distributed authentication architectures can be seen as an
evolutionary step, but also raises the issue always accompanied by an attempt to
mix disparate systems: Interoperability. What is the effect on performance of the
authentication process if an individual establishes his/her credential on one
2
system, and then authenticates him/her-self on a different system? This issue is
of relevance to all kinds of authentication mechanisms, and of particular
relevance to biometric recognition systems. The last decade has witnessed a
huge increase in deployment of biometric systems, and while most of these
systems have been single vendor, monolithic architectures the issue of
interoperability is bound to arise as distributed architectures become more
pervasive.
Fingerprint recognition systems are the most widely deployed biometric systems,
and most commercially deployed fingerprint systems are single vendor systems.
A recent market report published by International Biometric Group (IBG) (2007)
found that fingerprint recognition systems account for approximately 68% of the
biometric system deployments. The single point of interaction of a user with the
fingerprint system is during the acquisition stage, which has the maximum
probability of introducing inconsistencies in the biometric data. Fingerprint
sensors are based on a variety of different technologies like electrical, optical,
thermal etc. The physics behind these technologies introduces inconsistent
distortions and variations in the feature set of the captured image, which makes
the goal of interoperability even more challenging. There are obvious technical,
commercial and operational advantages to understanding these issues more indepth, in addition to making the objective of creating a globally interoperable
biometric system a more realistic one.
This dissertation is organized into five chapters. The first chapter discusses the
statement of the problem, the significance of the problem being addressed, and
rationale behind this research. The second chapter covers background material
and review of literature available in related areas. Selected works in the area of
fingerprint recognition, which are relevant to this dissertation, are summarized.
The third chapter outlines the methodology of research, the experimental setup
and the analysis framework to be followed during this study. The fourth chapter
3
contains the results and analysis of the procedures described in the third chapter.
The fifth chapter contains conclusions from the results analysis and
recommendations for future work.
1.2. Statement of Problem
The distortions and variations introduced when acquiring fingerprint images
propagate from the acquisition subsystem all the way to the matching subsystem.
These variations ultimately affect performance rates of the overall fingerprint
recognition system. Fingerprint images captured using the same sensor
technology during enrollment and recognition phases will introduce similar
distortions, thus making it easier to compensate for such distortions and reducing
its effect on the performance of the overall fingerprint recognition system.
However, when different fingerprint sensor technologies are used during
enrollment and recognition phases, an impact on performance is expected, but is
unpredictable.
Financial institutions provide a relevant example of the requirement for
interoperability of fingerprint sensors. Some institutions are starting to deploy
Automated Teller Machines (ATM) which use fingerprint recognition for
authenticating customers. Such a system can be designed to take advantage of
distributed acquisition architecture and use a centralized storage and matching
architecture as shown in Figure 1.1. Without proper understanding of how
fingerprints captured from different sensors affect the overall recognition rates, a
financial institution would be forced to deploy the same fingerprint sensor at all of
their ATM’s. An extraordinary level of confidence and trust in the fingerprint
sensor manufacturer would be required in order to choose just a single
manufacturer. The lack of choice could also be a hurdle to mass adoption of this
technology. If the sensor manufacturer was to stop supporting the particular
fingerprint sensor, a financial institution would be forced to replace all the
4
sensors and re-enroll all its clients. A financial institution would incur massive
capital and labor costs, which could be a deterrent to using this technology.
There is need to understand the effect of different fingerprints on recognition
rates not just from an algorithm advancement perspective, but also from a
technology usage perspective.
Figure 1.1 Example of an Interoperability Scenario.
1.3. Significance of Problem
One of the main components of minimizing error rates in a fingerprint recognition
system is the acquisition subsystem since it is the first point of contact between
the user and the system. Sensor variability results in moving the distribution of
genuine scores away from the origin, thereby increasing error rates and
negatively impacting the performance of the system (Wayman, 1997). Bolle and
Ratha (1998) describe a list of challenges of matching two fingerprints. Existence
of spurious features and missing features compared to the database template,
transformation or rotation of features, and elastic deformation of features are
several problems faced by fingerprint matchers. The acquisition subsystem is
responsible for introducing part or all of the variability. This issue is amplified
5
when a fingerprint recognition system is deployed in a distributed architecture
because of the use of different components during acquisition, signal processing,
and matching operations. A fingerprint recognition system deployed in a
distributed architecture can benefit from gaining a deeper insight into
interoperability of sensors and its effect on error rates.
Jain and Ross (2004) observed that Equal Error Rates (EER) for a fingerprint
dataset consisting of images acquired from two different sensors was 23.10%
whereas the EER for the two native datasets were 6.14% and 10.39%. In a
report submitted to U.S. Department of Agriculture, Ford and Sticha (1999)
outlined feasibility of biometric systems in reducing fraud in the food stamp
program and other welfare programs, and paid particular attention to fingerprint
recognition. In 2001 eight states in the U.S.A. required applicants for welfare in
at-least some counties to submit to fingerprint recognition (Dean, Salstrom, &
Wayman, 2001). These previously mentioned deployments are distributed
architectures which acquire fingerprints from various locations and have a higher
degree of centralization for its matching and storage functionalities. The ability of
a recognition system to integrate different fingerprint acquisition technologies
without a significant degradation of error rates has definite technological and
financial advantages. It not only allows the owners of the deployment to pick and
choose sensors that best fit its application, but also have the ability to switch a
few of the sensors without overhauling the entire deployment, or re-enrolling its
entire user-base. Campbell and Madden (2006) concluded that sensor effects on
performance are not trivial, and need to be investigated independently. A deeper
understanding of the effect of matching fingerprints acquired from different
sensors is imperative for mass adoption of this technology.
IBG (2007) has provided an analysis of adoption of biometric technologies and
applications, and a forecast of global revenues from 2007-2012 in their report.
They project annual revenues to grow from $3,012 million USD to approximately
6
$7,407 million USD. Commercial fingerprint recognition systems currently make
up approximately 25% of the entire biometrics market, and Automated
Fingerprint Identification Systems (AFIS) and other civilian fingerprint systems
make up approximately 33% of the entire biometrics market. Revenue generated
from the biometric market is projected to double in the next five years, and
fingerprint recognition systems are poised to generate a majority of it. Several
industry support and government enforced initiatives are expected to be prime
growth drivers for the biometrics industry but the issue of interoperability will have
to be tackled for a realization of this growth potential. An ever increasing
networked world is going to necessitate use of distributed architectures, and the
ability to integrate different fingerprint sensors in a heterogeneous system without
a significant impact on performance will be of critical importance.
Annual Biometric Industry Revenues
2007-2012
8000
6000
Revenue $m
USD
4000
2000
0
2007
2008
2009
2010
2011
2012
Year
Figure 1.2 Annual Industry Revenue Projections 2007-2012 (IBG, 2007).
7
Biometric Market by Technology
2007
AFIS/ LiveScan, 33.60%
Other
Modalities,
4.00%
Multiple
Biometric,
2.90%
Fingerprint,
25.30%
Vein
Recognition,
3.00%
Iris
Recognition,
5.10%
Voice
Recognition,
3.20%
Middleware,
5.40%
Face
Recognition,
12.90%
Hand
Geometry,
4.70%
Figure 1.3 Biometric Market Share by Technology (IBG, 2007).
1.4. Purpose of Study
Fingerprint recognition systems are primarily based on two different types of
matchers: minutiae based matchers, and pattern based matchers. The purpose
of this study was to examine the effect of sensor dependent variations and
distortions, and characteristics of the sensor on the interoperability matching
error rates of minutiae based fingerprint recognition systems. This study
achieved this aim by acquiring fingerprints from different acquisition technologies,
and examining error rates for native and interoperable fingerprint datasets. This
study did not attempt to isolate or examine specific variations and distortions in
different fingerprint acquisition technologies. The ultimate effect on matching
error rates of fingerprint datasets because of distortions introduced by different
technologies was the focal point of analysis. The outcome of such a study will be
useful in designing matching algorithms which are better at handling fingerprint
images from various sensors. The problem space affecting error rates of
fingerprint datasets is vast – every subsystem of a fingerprint recognition system
8
has an influence on performance of a fingerprint dataset. This study examined an
exclusive aspect of the problem space - the acquisition subsystem. The end
objective of this study was to provide greater insight into the effect of a fingerprint
dataset acquired from various sensors on performance measured in terms of
false non match rates (FNMR) and false match rates (FMR).
Figure 1.4 Purpose of Study.
1.5. Definitions
Biometric Algorithm represents the finite number of steps used by a biometric
engine to compute whether a biometric sample and template is a match (M1,
2004).
9
Biometrics is defined as automated recognition of individuals based on their
behavioral and biological characteristics (ISO/IEC JTC1 SC37 SD2 - Harmonized
Biometric Vocabulary, 2006).
Biometric Sample is the raw data representing a biometric characteristic of an
end-user as captured and processed by a biometric system (ISO/IEC JTC1 SC37
SD2 - Harmonized Biometric Vocabulary, 2006).
Biometric System is an automated system capable of the following (ISO/IEC
JTC1 SC37 SD2 - Harmonized Biometric Vocabulary, 2006):
1. capturing a biometric sample from the end user
2. extracting biometric data from the sample
3. comparing the biometric data to one or more reference templates
4. computing how well they match
5. indicating based on decision policy if identification or verification has been
achieved
Core is the turning point on the inner most ridge of a fingerprint (Amin & Yager,
2004).
Delta is the point on a ridge at or nearest to the point of divergence of two ridge
lines, and located at or directly in front of point of divergence (M1, 2004).
Equal Error Rate (EER) is the operational point where FNMR=FMR
Enrollment is the process of converting a captured biometric sample and the
subsequent preparation and storage of the biometric template representing the
individual’s identity (M1, 2004).
10
Environment is defined as the physical surroundings and conditions where
biometric capture occurs, and it also includes operational factors such as
operator skill and enrollee cooperation level (Biometric Sample Quality Standard
- Part 1: Framework, 2006).
False Accept Rate (FAR) is the expected number of transactions with wrongful
claims of identity that are incorrectly confirmed (Mansfield & Wayman, 2002).
False Reject Rate (FRR) is the expected number of transactions with truthful
claims of identity that are wrongly denied (Mansfield & Wayman, 2002).
False Match Rate (FMR) is the expected probability that a sample will be falsely
declared to match a single randomly selected non genuine template (Mansfield &
Wayman, 2002).
False Non Match Rate (FNMR) is the expected probability that a sample will be
falsely declared not to match a template from the same user supplying the
sample (Mansfield & Wayman, 2002).
Failure to Acquire (FTA) Rate is the expected rate of sample acquisitions which
cannot be processed.
Failure to Enroll (FTE) Rate is the expected proportion of subjects who cannot be
enrolled in the system.
Friction Ridge is the ridges present on skin of the fingers which make contact
with an incident surface under normal touch (M1, 2004).
Hybrid testing describes a methodology in which samples are collected in real
time and testing of the samples is performed in an off-line environment (Grother,
2006).
11
Identification is the one-to-many process of comparing a submitted biometric
sample against all of the biometric reference templates on file to determine
whether it matches any of the templates (M1, 2004).
Interoperable fingerprint dataset refers to the fingerprint dataset of 3 enrollment
fingerprint images and 3 testing fingerprint images collected from two different
sensors.
Live capture is the process of capturing a biometric sample by an interaction
between an end user and a biometric system (ISO/IEC JTC1 SC37 SD2 Harmonized Biometric Vocabulary, 2006).
Native fingerprint dataset refers to the fingerprint dataset comprised of 3
enrollment fingerprint images and 3 test fingerprint images collected from the
same sensor.
Performance is measured in terms of false non match rates and false match
rates which are fundamental error rates in offline testing (Grother et al., 2006).
Receiver Operating Characteristic (ROC) curves are a means of representing
results of performance of diagnostic, detection and pattern matching systems
(Mansfield & Wayman, 2002).
Template is the data which represents the biometric measurement of an enrollee
and is used by the biometric system for comparison against subsequently
submitted samples (ISO/IEC JTC1 SC37 SD2 - Harmonized Biometric
Vocabulary, 2006).
Valley is the area surrounding a friction ridge which does not make contact with
incident surface under normal touch (M1, 2004).
12
Verification is the process of comparing a submitted biometric sample against the
biometric reference template of a single enrollee whose identity is being claimed,
to determine whether is matches the enrollee’s template (M1, 2004).
1.5.1. Sensor Related Distortions and Variations
Fingerprint sensors are responsible for capturing the unprocessed representation
of ridge flow from the finger skin. Invariably, the capture process introduces
undesirable changes which make it different from the source of the sample.
These undesirable changes are called distortions. The causes of fingerprint
image distortions and variations can be categorized into the following:
1. Interaction type
2. Acquisition technology type
3. Sensor characteristics
The two most popular modes of interacting with fingerprint sensors are touch and
swipe. All sensors used in this study were one of these two types of interactions.
Both of these interaction types have a distinct effect on the image. Swipe sensors
utilize an image reconstruction technique which takes into account nonlinear
distortions between two consecutive frames (Zhang, Yang, & Wu, 2006). Touch
sensors are affected by elasticity of skin and amount of pressure placed on the
finger when it is placed on the sensor. Swipe sensors are not affected by residue
and latent prints, which can affect touch sensors. The artifacts introduced by
latent prints can distort the image captured from a touch sensor.
There are several different types of acquisition technologies used in fingerprint
sensors: optical, capacitive, thermal, ultrasonic, piezoelectric etc. The acquisition
technology introduces distortion because of the physics behind it. The three
types of acquisition technologies used in this study were optical, capacitive, and
thermal.
13
Most optical sensors are based on the phenomenon of frustrated total internal
reflection (FTIR) (O'Gorman & Xia, 2001). This technology uses a glass platen, a
light source and a Charged Coupled Device (CCD) or Complementary Metal
Oxide Semiconductor (S3OS) camera for constructing fingerprint images
(O'Gorman & Xia, 2001). Optical sensors introduce distortions which are
characteristic of its technology. The edges of fingerprint images captured using
optical sensors have a tendency of getting blurred due to the setup of the lenses.
Optical physics could potentially cause out of focus images which can be
attributed to the curvature of the lens. Sometimes residual incident light is
reflected from the ridges which can lead to a low contrast image (Secugen). A
phenomenon called Trapezoidal Distortion is also noticed in fingerprint images
captured from optical sensors due to the unequal optical paths between each
point of the fingerprint and the image focusing lens (Igaki, Eguchi, Yamagishi,
Ikeda, & Inagaki, 1992). The level of contrast in resulting fingerprint images is
affected by the acquisition technology. Optical sensors tend to introduce grey
areas in the image due to residual light getting scattered from the ridge and not
reflecting completely.
Capacitance sensors do not produce geometric distortions, but they are prone to
introduce distortions due to the electrical nature of the capture technology.
Electrostatic discharge can affect the resulting image since the conductive plates
are sensitive to it. Capacitance sensors can also be affected from the 60Hz
power line and electrical noise from within the sensor (2006).
Thermal sensors also do not introduce geometric distortions. They work on the
principle of measuring difference in heat flux, which makes the sensor
susceptible to humid and warm conditions. Since this type of technology
measure the difference in heat flux, the finger surface has to be swiped across
the sensor and introduces distortions typical of swipe technology.
14
Sensor characteristics like sensing surface and resolution are also responsible
for introducing distortions. The horizontal and vertical resolutions are responsible
for determining the distance observed between pairs of minutiae points. Sensors
with different resolutions will provide images which show a different Euclidean
distance for the same pair of minutiae points. The area of the sensing surface will
determine the level of overlap in the same fingerprint image captured on a
different sensor. It is more likely for two different sensors with a similar sensing
areas to capture the same fingerprint image with a higher level of overlap
compared to sensors with different sensing area.
1.6. Assumptions
•
All participants were willing participants in the study, and thus were treated
as cooperative participants and provided samples in genuine and good
faith.
•
The integrity of fingerprint sensors was maintained throughout the study.
•
The fingerprint sensors, minutiae extractors and minutiae matchers were
either open source or commercially available. Their integrity and
performance were assumed to be consistent with their product
specifications.
1.7. Delimitations
•
The participants in the study were recruited from Purdue University.
•
All data collection was performed in a single visit.
•
The participants in the study were allowed to adjust the height of the chair
and positioning of the fingerprint sensor to suit their comfort needs.
•
Mistakes in placement of finger on sensor invalidated the fingerprint
sample provided by the participant.
15
•
All fingerprint sensors were peripheral devices connected using USB port.
•
An ultrasonic fingerprint sensor was not used.
•
Environment conditions were controlled throughout the data collection
process.
•
Not all participants were familiar with fingerprint recognition technology. All
participants had to complete a practice session to familiarize themselves
with the process of providing fingerprint images.
•
Two different fingerprint feature extractors and fingerprint feature
matchers were used: VeriFinger 5.0 extractor and matcher, and MINDTCT
and BOZORTH3. VeriFinger 5.0 extractor and matcher create one
subsystem and MINDTCT and BOZORTH3 create another subsystem.
This study did not compare interoperability of the different subsystems.
1.8. Limitations
•
The experience of participants with fingerprint sensors participating in the
study was different. Previous studies have observed that prior experience
and habituation of users with fingerprint devices has an effect on
performance rates (Thieme, 2003).
1.9. Summary
This chapter introduced the problem in global terms by giving a description of the
problem space and its significance to overall advancement of fingerprint
recognition. An overall frame of reference is created for the reader which will
assist in highlighting the focus areas of this study.
16
CHAPTER 2. REVIEW OF LITERATURE
2.1. Introduction
This chapter outlined several studies and experiments related to the impact of
sensor specific issues on fingerprint recognition systems. Previous studies were
analyzed to identify and justify the design of experiment. The statistical analysis
methodology used in this dissertation was formulated based on previous
research analysis methodologies. This chapter also presented a historical and
evolutionary view of fingerprint recognition systems.
2.2. History of Biometrics
Archaeological evidence has been found indicating fingerprints were used as a
form of authentication dating back to the 7000 to 6000 B.C. (O'Gorman, 1999).
Clay pottery from these times used a fingerprint impression to mark the identity of
the potter who had made it. Bricks used in the ancient city of Jericho have been
discovered with impression of fingerprints, and its most probable use was to
recognize the mason. Likewise, in ancient Egypt allowance claimants were
identified by checking against a record which contained the name, age, place of
origin, and other relatively unique physical and behavioral characteristics
(Ashbourn, 2000).
2.2.1. Studies Related to Finger Anatomy and Ridge Structure
The earliest scientific fingerprint studies aimed at studying the anatomy and
function of papillary ridges on the surface of the finger. Two microscopists,
Govard Bidloo and Macello Malphighi published their findings and conclusions on
17
papillary ridges in 1685 and 1687 respectively (Cole, 2001). The main point of
discussion in those publications was if papillary ridges were organs of touch or if
they facilitated sweating. Anatomist J.C. Mayer (1788) made a claim in his
publication that the arrangement of skin ridges could never duplicated in two
persons. This was among the first recorded claim about uniqueness of
fingerprints. In 1823 Czech physician Jan Purkyne classified papillary ridges on
human fingerprints into nine categories in an effort to establish a connection
between vision and touch (Hutchings, 1997). This was the first attempt at
classifying fingerprint patterns and laid the foundation for future fingerprint
identification systems. Although Purkyne was a trained physician, he was also a
student of philosophy and believed that every natural object is identical to itself
and thus claimed that no two fingerprint patterns are identical.
2.2.2. Early Experiments with Fingerprint Identification: Criminalistics
William Herschel and Henry Faulds were the first to apply fingerprints for
verifying the claim of an individual’s identity by checking it against a catalogue of
recorded fingerprints. In 1858 Willliam Herschel asked a road contractor to
impress his hanS4rint in ink on a deed with the intention of adding the element of
non-repudiation to the contract. Upon further investigation of the details of the
hanS4rint he believed that he had found a way of verifying the identity of every
man, and tried to institute it as an identification method in a local prison to identify
the inmates, but his ideas did not impress his superiors and his ideas were never
tested. The impracticality of comparing a visual representation of a fingerprint
with all the catalogued fingerprints was its primary drawback. In 1880 Henry
Faulds observed that fingerprint patterns could be used to identify criminals.
Faulds devised an alphabet classification scheme so that a person’s set of ten
prints could be represented by a word. These words could be catalogued in a
dictionary which would then be used to search for fingerprints. Faulds published
his observations and methodology in a letter to Nature in 1880 and also
approached Scotland Yard with his ideas. Although he provided an improvement
18
over previous methods with the use of a cataloguing system, Scotland Yard
declined to follow up on his technique.
2.2.3. Galton’s Fingerprint Experiments
In 1880 Francis Galton followed up on Fauld’s experiment on using fingerprints to
identify individuals. Galton started off by devising a classification system which
built on Purkyne’s 9 pattern types, but he found that they were not discriminative
enough. There were fingerprints with patterns which blurred the division between
the 9 types. Eventually Galton came up with 60 different patterns which he
believed to be discriminative and inclusive but then realized the impracticality of
such a system. He reconsidered his efforts and reclassified all patterns into three
categories: “arches”, “loops” and “whorls”. Galton recognized the key to
fingerprint classification lay in grouping, not in differentiating. The empirical
results from Galton’s experiments showed that the patterns were not uniformly
distributed. The loop pattern was the most common pattern, so Galton divided it
into inner loop pattern and outer loop pattern. According to Galton’s classification
inner loops opened towards the thumb and outer loops opened towards the little
finger. For purposes of classification Galton suggested classifying all ten fingers
based on the four pattern types and using that combination of patterns to
catalogue the individual. Galton demonstrated his classification scheme to the
British Home Office in 1893 but non uniform distribution of fingerprint patterns
posed a challenge of representing large collections, which dissuaded them from
adopting Galton’s scheme.
In the course of his experiments he noticed features of ridges where the papillary
ridge ended, split into two, or rejoined from two ridges into one. He went on to
propose that two fingerprints could be matched by comparing these features that
he called “minutiae”. Although theoretically this method had merit, Galton
believed this method to be infeasible in practical terms. The number of minutiae
that needed to match correctly in order to make a confident claim of a match was
19
a stumbling block for Galton’s minutiae matching method. This discovery would
not be used to its fullest strength until advent of automated minutiae matchers in
1960’s.
Figure 2.1 Examples of Galton features.
2.2.4. Henry’s Fingerprint Experiments
Edward Henry was a colonial police officer in India interested in using fingerprints
to identify criminals since he believed the Bertillon system based on
anthropometrics was inadequate (Cole, 2004). Henry and his assistants set
about creating a feasible and practical fingerprint classification scheme and
devised a solution using ridge counting and ridge tracing. They subdivided loops
by counting the number of ridges between the delta and core. They subdivided
whorls by tracing the ridge from the delta and determining whether it passed
inside, outside or met the core. They also added a fourth group called
composites to the original three classification groups. The concept of ridge
tracing and ridge counting was crucial in making it workable. Henry and his
assistants, Haque and Bose, built upon the concept of ridge tracing and ridge
counting and created a heuristic based fingerprint classification methodology
which was capable of accommodating up to 100,000 prints (Haylock, 1979).
20
Henry introduced this system in his jurisdiction in 1895 and by 1897 it was
adopted all over India. Henry published his system in a publication called “The
Classification and Uses of Finger Prints” in 1900 and by 1902 Scotland Yard had
fully adopted fingerprinting for purposes of identification.
2.2.5. Fingerprint Recognition in South America, Europe
Parallel development of fingerprint identification and verification techniques was
being pursued in Argentina and France in late 19th century and early 20th century.
Juan Vucetich published “Dactiloscopia Comprada” in 1904 which is one of the
earliest notable works on fingerprinting (Polson, 1951). Vucetich came across
Galton’s lecture on “Patterns in Thumb and Finger Marks” which motivated him to
apply fingerprinting to the problem of reliable identification. Vucetich developed a
10 digit system based on Galton’s classification. He used the four basic pattern
types of arch, left loop, right loop and whorl. In addition, he used a secondary
classification of five sub-types: loop with plain pattern, loop with adhering ridges,
internal loop approximating a central pocket, external loop approximating a
central pocket, and irregular loops that did not include any of the previous types
(Cole, 2004). In 1905 an autonomous finger print bureau was created in
Argentina based on Vucetich’s fingerprint identification system.
In 1891 in France, Forgeot described a technique of development of latent
fingerprints based on previous work on micro-chemical tests for analysis of
sweat. Locard of Lyons is credited with inventing poroscopy when he
demonstrated that pores of skin are relatively unique in shape, size and position
in each individual (Polson, 1951). Locard applied his method in the BoudetSimonin case which led to a conviction of burglary. Bertillon, who is considered
widely to be the inventor of anthropometry, also used bloody fingerprints in case
of Henri Scheffer to prove the identity of the murderer.
21
2.3. Automated Fingerprint Recognition
Attempts to automate the process of identification can be found dating back to
1920. IBM punch card sorters and other automated data processing technologies
were employed to alleviate the problem of handling large number of cards. In
1919 the California State Bureau of Identification introduced a mechanized punch
card system called the Robinson F-index to assist in storing and retrieval of
information. In 1934 the Federal Bureau of Investigation (FBI) began using an
IBM card sorter to search coded fingerprint classifications, and in 1937 the New
York State Division of Criminal Identification started using an IBM card sorter
which could sort 420 cards per minute (Cole, 2001). These approaches solved
part of the problem, but human examiners were still required to inspect
fingerprints. The need to automate the entire matching process was widely
acknowledged, but technology constraints remained a barrier to progress. The
first experiments with optical recognition of fingerprint images began in the
1960’s (Allen, Prabhakar, & Sankar, 2005). In 1963 Joseph Wegstein and
Raymond Moore began work on an automated fingerprint identification system
under the auspices of FBI and National Bureau of Standards (Reed, 1981). In
1972 the FBI installed a system with a fingerprint scanner built by Cornell
Aeronautical Laboratory and a prototype fingerprint reader system built by North
American Aviation (Cole, 2001). With considerable progress achieved on
digitizing and automating the matching process, the FBI started scanning all their
fingerprint records for persons born after January 1, 1929 and by 1980 they had
a databank of 14.3 million records. Throughout the 1980’s various city police
departments and state criminal justice bureaus in the U.S.A. started deploying
Automated Fingerprint Identification Systems (AFIS) confirming the maturity of
the technology.
2.4. General Biometric Recognition System
A generic biometric authentication system can be divided into five subsystems
based on its functionality: data collection, transmission, signal processing,
22
decision and data storage (A. Jain, Maltoni, Maio, & Wayman, 2005). Wayman
(1997) described a general biometric model in “A Generalized Biometric
Identification System Model” The data collection subsystem samples the raw
biometric representation and the sensor converts it into an electronic
representation. The transmission subsystem is responsible for transportation of
the electronic representation of the biometric to the signal processing subsystem.
Data compression and adjustment for noise are the main tasks of this
subsystem. The signal processing subsystem takes the biometric signal and
converts it into a feature vector. The tasks of this subsystem differ according to
the modality being processed, but quality assessment and feature extraction is
performed by this subsystem. The storage subsystem is responsible for storing
the enrolled templates, and depending on the policy governing a particular
biometric system the raw signal might also be stored. Binning and classification
of templates according to predefined criteria are also employed within this
subsystem in order to increase efficiency of the matching process. Inputs to the
decision subsystem are measures resulting from comparison between the
feature vector of the input sample and the enrolled templates. Using the system
decision policy, a match or a non-match is declared.
23
Signal
Data Collection
Processing
Biometric
Decision
Pattern Matching
Presentation
Quality Control
Sensor
Feature
Decision
Extraction
Storage
Transmission
Database
Expansion
Compression
Image
Transmission
Storage
Figure 2.2 The General Biometric System (Wayman, 1997)
R. M. Bolle, Connell, Pankanti, Ratha and Senior (2004b) described a biometric
authentication system as a pattern recognition system comprised of biometric
readers, feature extractors, and feature matchers. They categorized the
authentication system based on the processes of enrollment and authentication.
In their model, the enrollment subsystem captures the biometric measurements,
extracts the relevant features and then stores them in a database. Depending on
system policy, ancillary information like identification number can also be stored
with the biometric template. The authentication subsystem is concerned with
recognizing a individual once he/she has been enrolled in the biometric system.
24
Figure 2.3 Biometric Authentication System (R. M. Bolle, Connell, Pankanti,
Ratha, & Senior, 2004b)
Wayman et al. (2005) use the functionality of each subsystem, and process flow
as a means of describing a general biometric model, and Bolle et al. (2004) use
the different stages of the authentication process as a means of describing a
general biometric model. The differences in description of the two models
highlighted the possibilities for the underlying system architectures for biometric
systems.
A detailed model was developed within ISO/IEC JTC1 SC37 which described the
general biometric system in terms of the functional components of the system
and process flow in the system. A notable deviation from the previously
described biometric system model was removal of the transmission subsystem
as a specific part of the biometric model. In a distributed and networked
authentication system, transmission is a general function of the overall system
instead of a function specific to the biometric system.
25
Figure 2.4 ISO/IEC JTC1 SC37 Conceptual Biometric Model (ISO)
2.4.1. Fingerprint Recognition Model
A distributed authentication architecture is an fundamental reason for focusing on
the issues of interoperability. A distributed fingerprint recognition system can be
designed in a multitude of ways, and a list of possible architectural configurations
can be considered based on acquisition location of fingerprint sample, storage
location of the fingerprint template, and matching location of input sample and
fingerprint template. Table 2.1 shows the four possible processing locations.
The server is defined as a centrally located computer which is at a different
physical location than the requesting client. The local workstation is where a user
initiates interaction with the fingerprint recognition system. The peripheral device
can also be connected with the local workstation using input/output ports, or an
embedded device. Physical token refers to smartcards, PS3CIA cards and other
26
small scale devices that could support any of the required processes. Figure 2.5
shows a list of all possible architectural configurations. There are 43 = 64
configurations possible, but not all of them are feasible in a practical
implementation. The purpose of depicting these possible configurations is to give
an idea about the complexity of designing a distributed architecture.
Table 2.1 Possible Acquisition/Storage/Matching Location
Server
Local Workstation
Peripheral Device
Physical Token
Figure 2.5 Possible Distributed Architectures
Several live implementations use one of the possible configurations as their
distributed architecture design. Digital Persona U.are.U Pro for Active Directory
replaces passwords with fingerprint recognition for authentication used in
27
enterprise network applications (DigitalPersona, 2006). This product allows
organizations to centralize the network authority processes and is an example of
distributed acquisition and centralized storage/matching architecture. The Clear
Registered Traveler program provides registered members with a smart card that
stores fingerprint, face, and iris templates for card’s owner ("Clear Registered
Traveler", 2006). For verification purposes the registered individual steps up to a
kiosk, provides their fingerprint, face, and iris samples which are matched with
the templates on the card. In this scenario, the matching happens on the
machine in the kiosk, and this is an example of architecture that uses a physical
token to store, and a local machine to perform the matching operation. In this
architecture, the storage and matching operations are performed on separate
devices. The increasing use of distributed architectures brings into focus the
issues of interoperability. The following section describes the history and
development of acquisition technologies used in fingerprint recognition systems.
2.4.2. Fingerprint Acquisition Technologies
The inked method of fingerprint acquisition was the preferred means of data
capture from the 1870’s when fingerprints were first considered for law
enforcement purposes until the 1960’s when the first generation AFIS was
developed. Advances in storage of digital information and automated recognition
of fingerprints led to innovations in fingerprint acquisition technologies. Rapid
advances were made in this field and the following is a brief listing of the family of
imaging technologies:
•
Optical
•
Electrical
•
Thermal
•
Ultrasonic
•
Piezoelectric
28
Optical, electrical and thermal are the most popular types of fingerprint
acquisition and imaging technologies which will be discussed more in-depth in
the following section. The proportion of ultrasonic and piezoelectric fingerprint
acquisition devices is relatively small compared to the other three types of
sensors and were not considered for this research, but it still does deserved
mention because it is a feasible method of capturing fingerprints.
2.4.2.1. Optical Sensors
Optical fingerprint capture devices have the longest history of all automated
fingerprint image acquisition devices (O'Gorman & Xia, 2001). The earliest
methods of optical capture involved using a camera and taking direct images of
fingerprints. This method had several issues - the most prominent one was that
fingerprint ridges and valleys were not differentiated by color and a shadow
needed to be introduced to differentiate between ridges and valleys (Setlak,
2004). Using the principal of frustrated total internal reflection (FTIR) was a major
step forward for optical sensors.
When light is incident on the interface of two media, it is partly transmitted
into the second medium and partly reflected into the first. If however, the
index of refraction of medium 1 is greater than the index of medium 2 and
the angle of incidence exceeds the critical angle, total internal reflection
occurs. The incident light is reflected completely back into medium 1
(Hawley, Roy, Yu, & Zhu, 1986).
The refractive index is different for ridges and valleys when a finger touches the
platen of the optical sensor. Due to this phenomenon, light incident on valleys
gets totally reflected and the light incident on ridges is not reflected which results
in the ridges appearing dark in the final image (O'Gorman & Xia, 2001).
29
Figure 2.6 Optical Fingerprint Sensor
Figure 2.7 Fingerprint Image from Optical Fingerprint Sensor
As shown in Figure 2.6 this technology focuses light incident on the ridges and
valleys onto a CCD/S3OS camera which captures the image. Older generation
optical sensors used CCD cameras, but newer generation optical sensors used
S3OS cameras. CCD cameras are usually analog which requires additional
hardware to convert the analog signal to digital signal. S3OS cameras
incorporate an analog-digital conversion process thus simplifying the cost and
complexity of the system (O'Gorman & Xia, 2001).
30
Along with S3OS cameras, multiple mirrors and sheet prisms are also used to
reduce size of sensors. The reason for introducing reflective surfaces within the
optical sensor is to reduce the total optical length between the finger surface and
the camera. The optical length can be modified by changing the lens focus and
width of the camera array. Using multiple mirrors and sheet prisms is a method
used to reduce the optical length, thereby reducing the size of the sensor.
Optical sensors are not susceptible to electrostatic discharge which can disrupt
the image capture process in solid state capacitive sensors. If a direct light
source is pointed at the sensor it could have an effect on the fingerprint image.
Optical fingerprint sensors have a large platen area which allows it capture a
large portion of the fingerprint thereby providing a larger area of the fingerprint in
the image. The large size of optical fingerprint sensors is a disadvantage since it
makes them impractical for use in small scale devices like cell phones. Optical
fingerprint sensors are also affected by residue on the platen which could lead to
an imperfect representation of the fingerprint.
2.4.2.2. Solid State Capacitive Sensors
Capacitive sensors work on premise that skin on finger surface is an equipotential surface – constancy of potential is a requirement to obtain
representative fingerprint images (Lim & Moghavvemi, 2000). In the late 1960’s
several inventors proposed methods of capturing fingerprint images using two
dimensional conductive arrays. Rapid advances were made during the 1980’s
that led to new approaches to capacitive sensing technologies. But these early
concepts all had a common drawback; they did not use integrated circuits that
could take advantage of signal amplification and scanning. In the late 1980’s
researchers working with memory devices discovered that a finger place on a
memory array caused data errors that resembled the spatial patterns of the
fingerprint (Setlak, 2004). These memory chips used periodically refreshed
charge stored on a capacitor in the memory cell and the difference in
31
capacitance between ridges and valleys on skin of the finger caused the bits to
flip. This phenomenon was refined and has been present in commercially
available capacitive sensors available since mid 1990’s.
Capacitive sensors are constructed using a two-dimensional array of conductive
plates. The finger is placed on a surface above the array so that the electrical
capacitance of these plates is affected. The sensor plates under the ridge will
have a larger capacitance than the sensor plates beneath the valley. This is
because air has lower permittivity than skin, which leads to an increased
capacitance in plates under the skin.
Capacitance
Plates
Electrical
Lack of
Charge
Charge
Figure 2.8 Capacitive Fingerprint Sensor
Capacitive fingerprint sensors can be classified into two classes: single plate
capacitive sensors and double plate capacitive sensors. In a single plate
capacitive sensor each pixel has at-least a capacitive sensor, a sensor circuit
and additionally can have a logic circuit (Machida, Morimura, Okazaki, &
Shigematsu, 2004). Each pixel is charged separately and measured separately in
order to generate a complete fingerprint image. In double plate capacitive
sensors two adjacent conductive plates correspond to one pixel. The capacitance
differential between two adjacent plates is used to generate a pixel value.
32
Figure 2.9 Fingerprint Image from Capacitive Fingerprint Sensor
Capacitive sensors have a relatively smaller sensing area, which makes them
suitable for use in small scale devices. The smaller sensing area results in a
smaller image of the fingerprint which can be a disadvantage if the users are not
consistent with their finger placement. Capacitive sensors are susceptible to
electrostatic discharge, and to moisture content of the finger. They are not
affected by light reflecting on the sensor surface.
2.4.2.3. Thermal Sensors
Sensors were designed which use thermal energy flux to capture fingerprints.
When a ridge is in contact with a sensor surface of different temperature, heat
flows between the ridge and sensor surface. The sensor surface is made up of
an array of micro-heater elements, and a cavity is formed under between the
valley of the fingerprint surface and heater element (J. Han, Kadowaki, Sato, &
Shikida, 1999). Since the valley is not in contact with the sensor surface there is
no heat flow between the valley and sensor surface. The heat flux is measured
and converted into a digital representation of the fingerprint surface.
33
Heat
Flow
Heat
Flow
Air
Sensing
Surface
Figure 2.10 Thermal Fingerprint Sensor
When surfaces with different temperatures come in contact, they attempt to
reach equilibrium as quickly as possible. Because of this phenomenon heat flux
decays very rapidly when the finger is held is a static position. Thermal sensors
are designed such that a finger has to be moved over the sensor surface in a
constant motion in order to maintain an acceptable level of heat flux.
Figure 2.11 Fingerprint Image from Thermal Fingerprint Sensor
34
The thermal nature of the sensing technology makes it sensitive to environmental
temperature conditions which is one of its major drawbacks. Thermal sensors are
less influenced by moisture content of the finger (Adhami & Meenen, 2001).
Thermal sensors are also not affected by light reflecting on the sensor surface.
The small size of thermal sensors makes it a suitable choice for small scale
devices.
2.4.3. Fingerprint Image Quality Assessment
Automated and consistent quality assessment of input samples is an important
component of any biometric system. The ability of a system to detect and handle
samples of varied quality levels is a significant contributor to performance of a
biometric recognition system. The same holds true for fingerprint recognition
systems. Fingerprint sample quality assessment is a topic of great interest as it
can be used to ensure samples of appropriate quality are used for the matching
process. The term quality is used in three different contexts as it relates to
biometric sample quality (ISO, 2006):
1. Fidelity: reflects the accuracy of a sample’s representation of the original
source.
2. Character: reflects the expression of inherent features of the source.
3. Utility: reflects the observed or predicted positive or negative contribution
of the biometric sample to overall performance of a biometric system.
Quality assessment algorithms use fidelity, character, utility or a combination of
the three to compute a quality score. The following sections explain different
types of fingerprint quality assessment algorithms and related research studies.
2.4.3.1. Types of Fingerprint Image Quality Assessment Algorithms
Aguilar, Fernandez and Garcia (2005) categorized existing image quality
assessment algorithms into four broad categories:
35
1. Based on local features.
2. Based on global features.
3. Based on classifiers.
4. Hybrid algorithms based on local and global features.
They categorized algorithms which subdivided the fingerprint image into blocks,
and computed quality scores for each block as local feature quality algorithms.
This type of analysis takes into account specific local features. They categorized
algorithms which search for abrupt changes in ridge orientation as global feature
quality assessment algorithms. These algorithms tend to use 2-D discrete Fourier
transform and energy concentration analysis of global structure to assess image
quality of fingerprints. They third category of quality assessment algorithms were
based on the premise that a quality measure should define a degree of
separation between match and non-match distributions of a fingerprint. Using a
relatively large dataset, classifiers can be trained using degree of separation as a
response variable based on a vector of predictors, and then map the degree of
separation to a quality index. They categorized hybrid algorithms as the ones
which used an aggregation of local and global feature analysis to compute a
quality index. The next section describes research related to the four categories
of image quality assessment algorithms.
2.4.3.2. Experiments Related to Image Quality Assessment
Jain, Chen and Dass (2005) proposed a fingerprint image quality score
computation which used both global and local quality scores. Energy
concentration in frequency domain was used as a global feature. Their analysis
as based on the premise that good quality images had a more peaked energy
distribution while poor ones had a more diffused distribution and this property
was used to compute the global image quality score. The local image analysis
was performed by dividing the image into sub-blocks and computing the clarity of
local ridge-valley orientation in each sub-block. A single quality score was
36
computed as a weighted average of all the sub-blocks. These two methods were
tested for improving fingerprint matcher performance. A decrease of 1.94% in
EER was observed when images with the lowest 10% quality scores were
pruned from the FVC 2002 DB3 database.
Jian, Lim and Yau (2002) described an image quality assessment method which
used three different local and global features of the fingerprint. The first factor
was a local feature. The strength and direction of ridge orientation information
was computed since it determined contrast of the image and clarity of ridges. The
second factor was a global feature. A gray-level analysis was performed to
determine clarity of ridge valley structure. The third factor was an assessment of
continuity of ridges which was also a global feature. The local and global quality
scores were used to compute a final quality score. They also defined a quality
benchmark, which was a function of minimal area of the fingerprint image
required for authentication, number of falsely detected minutiae, number of
correctly detected minutiae, and number of undetected minutiae. Regression
analysis was performed by plotting the image quality score against the quality
benchmark score and the model had a R2 value of 0.76. The authors concluded
their method could be used reliably for image quality assessment.
Abdurrachim, Kuneida, Li and Qi (2005) described a hybrid method for fingerprint
image quality calculation which used seven indices. These seven indices were
computed from local and global feature analysis. A Gabor feature index,
smudginess and dryness index, foreground area index, central position of
foreground index, minutiae count index, and singular point index were used as
input parameters for a quality function. The weighting for each index was
computed using an overlapping area methodology and linear regression
methodology. Their results showed that their hybrid quality measure decreased
EER by 12% - 34% when 10% of the images with lowest quality score were
pruned from the database.
37
Jiang (2005) described a computationally efficient method for computing image
quality using local features. A two stage averaging framework was used to take
advantage of linear and normalized vector averaging. In the first stage,
orientation and anisotropy estimation was performed using principal component
analysis in order to determine local dominant orientation of the fingerprint image.
Anisotropy is defined as the property of being directionally dependent. The
modulus of the orientation vector was set to be its anisotropy estimate, whereas
previous methodologies had used unity to set the modulus of the orientation
vector. An improvement in EER was observed when compared to other methods
of averaging the gradient orientation vectors.
Tabassi and Wilson (2005) proposed a novel methodology for assessing
fingerprint image quality. They defined fingerprint image quality as a predictor of
matcher performance before the matching operation is performed. They
computed a normalized match score for every fingerprint in a database, and use
quality scores of 11 different local features of a fingerprint as a non-linear
combination to predict the normalized match score. They used a 3 layer feedforward artificial neural network as their non-linear classification method. 5,244
images were used for training the network, and 234,700 images were used for
testing the network. Their methodology divided quality into 5 levels, and matching
performance of images at those 5 levels was compared. Their results showed
that true accept rates (TAR) decreased and false accept rates (FAR) increased
as quality of the fingerprints decreased. The [FAR,TAR] for images at quality
level 1 (excellent quality) was [0.0096, 0.999] and for images at quality level 5
(poor quality) was [0.075,0.889].
Bigun, Fronthaler and Kollreider (2006) proposed a method to assess quality of
an image using local features. They used the orientation tensor to draw
conclusions about the quality of the image. The orientation tensor holds edge
information, and this idea was used to determine if information was structured.
38
Their method decomposed the orientation tensor into symmetry representations
and the symmetries were related to particular definition of quality. In order to test
their quality assessment algorithm they used the QMCYT fingerprint database
comprised of 750 unique fingers. Their proposed quality assessment algorithm
was applied to all the fingerprints, and the fingerprints were split into 5 equally
sized partitions of increasing quality. Each partition was tested on two different
fingerprint matching algorithms, and a decrease in equal error rate was observed
for each increasing quality partition. A reduction in EER by approximately 3%
was observed between the worst and best quality partitions.
Alonso-Fernandez et al. (2006) performed a study to examine the performance of
individual users under varying image quality conditions for a fingerprint database
collected using three different fingerprint sensors. One of the main aims of their
study was to investigate the effects of image quality in the performance of
individual users. They proposed using a score normalization scheme adapted to
the quality score of individual users. They collected 12 fingerprint images from 6
fingers of 123 participants on 3 different fingerprint sensors, which gave them a
total of 26,568 fingerprint images. Image quality scores and matching scores
were computed using NFIQ and BOZORTH3 respectively, both available from
NIST. On performing a correlation test for verification EER threshold and quality
scores they noticed a strong positive correlation between the two variables. They
formulated a score normalization technique which incorporated the quality score
of the fingerprint image and the verification scores in order to compute a new
normalized matching score. They observed an EER reduction of 17% for one of a
dataset of fingerprints collected using an optical sensor. They also observed that
the score normalization technique improved EER for datasets which contained
low quality fingerprints.
39
2.4.4. Fingerprint Feature Extraction
Most fingerprint matchers use minutiae details, ridge pattern details, or a
combination of both to perform the matching operation. Minutiae based matchers
extract minutiae details from the spatial domain fingerprint image and create a
minutiae map of these features (Eshera & Shen, 2004). The transformation of
spatial domain details to a minutiae map can be broken down into the following
steps: pre-processing, direction calculation, image enhancement, singular point
detection, and minutiae extraction(L. C. Jain, Erol, & Halici, 1999).
Direction calculation involves creating a direction map of the spatial domain
details. A true direction map will contain orientation of every pixel of the ridge
line, but this is a computationally expensive operation. In order to capture the
direction details in a manner that is efficient, the fingerprint image is divided into
blocks, and orientation is assigned to each block. Various algorithms have been
proposed which use either overlapping blocks or completely separate blocks.
Enhancement of a fingerprint image can be performed using a low-pass filter to
eliminate noise. Filters have been proposed that use average ridge width and
valley width as parameters. Since enhancement can have a significant impact on
performance of a fingerprint matcher, a lot of work has been done in this area to
explore different methodologies for enhancement. Binarization and
skeletonization of fingerprint images involves eroding the ridge lines until their
width is just a single pixel. The process of binarization and skeletonization
reduces the amount of information needed to process for minutiae extraction
(Jain et al., 1999). The core and delta are categorized as singularity points.
Calculation of the rate of change of orientation on a ridge line has been used to
detect the singularity points, also known as Poincare index. Core and delta points
on a fingerprint image will typically have a Poincare index between -1/2 and +1/2,
whereas ordinary points will have a Poincare index of 0 (Muller, 2001). Any
discontinuity in the ridgeline flow is categorized as a minutiae point, and different
methodologies for extracting these minutiae points have been proposed. Most of
40
these algorithms will do some sort of ridgeline tracing and mark out all the
discontinuities, and use the direction map to categorize their type. The next
section discusses feature extraction methodologies in detail. The following
section describes several experiments related to minutiae extraction.
2.4.4.1. Experiments Related to Minutiae Extraction
The FBI minutiae reader uses the binarization method as part of its minutiae
extraction process. A composite approach based on local thresholding and a slit
comparison formula was used to compare alignment of pixels along the dominant
direction (Stock & Swonger, 1969).
Boashash, Deriche and Kasaei (1997) described a feature enhancement and
feature extraction algorithm that used foreground/background segmentation
which leads to a more precise ridge extraction process. The segmentation was
based on the premise that in a given block, noisy regions had no dominant
direction, which lead to foreground regions exhibiting a high variance in direction
orthogonal to the orientation of the pattern, and a low variance along its dominant
ridge direction (Mehtre, 1993). The ridge extraction process looked at each pixel
in each foreground block with a dominant direction and determined if the pixel
was part of the ridge. The obtained ridge image was then thinned to width of one
pixel. Boashash et. al (1997) designed a minutiae extraction process which
calculated the Crossing Number (CN) at point a point P which was expressed as
(pp.5):
8
CN = 0.5 ∑ | Pi − Pi + 1 | P9 = P1
i =1
Eq. 2.1
where Pi was pixel value in 3 X 3 neighborhood of P as shown in Figure 2.12
41
P4
P3
P2
P5
P
P1
P6
P7
P8
Figure 2.12 3X3 matrix of P (Boashash et. al, 1997)
Depending on the calculated CN for a point, it could be categorized as an
isolated point, end point, continuing point, bifurcation point or a crossing point.
The x-coordinate, y-coordinate, and local ridge direction θ were extracted for
every end point and bifurcation point.
Maio and Maltoni (1997) presented a direct gray scale minutiae detection
approach without binarization and thinning. Their methodology followed ridge
lines on a gray scale image according to the fingerprint directional image. A set
of starting points was determined, and for each starting point the algorithm kept
following the ridge lines until they intersected or terminated. They defined the
ridge line as a set of points which were local maxima along one direction. The
extraction algorithm located at each step a local maximum relative to a section
orthogonal to the ridge direction. The consecutive local maxima points were
connected to form an approximation of a ridge line. Once all ridge lines were
detected, the minutiae points were detected by following the ridge lines along
both directions of the ridge line. Their performance results showed a high
proportion of errors caused by end point minutiae being detected as bifurcation
minutiae points.
Foo, Ngo and Sagar (1995) described a fuzzy logic minutiae extraction which
used a set of small windows to scan over the entire fingerprint image. A set of
fuzzy logic based rules based on pixel intensities were used to determine if the
42
window contains minutiae points. “Dark” and “bright” labels were used to
describe gray pixel values. The membership functions mapped the gray pixel
values against a degree of membership towards dark or bright categories. Their
approach used a combination of six sub-windows to model the possible types
and orientation of ridge endings and bifurcations.
Chan, Huang, and Liu (2000) proposed a novel methodology for extracting
minutiae from gray level fingerprint images. Their methodology used the
relationship between ridges and valleys to determine minutiae locations in a
fingerprint. The relationship between ridges and valleys are affected by minutiae
due to the discontinuities that minutiae introduce. A ridge and its two
surrounding valleys ere traced, and if the two valleys met at a point then the
heuristic decided a minutiae ending was detected. The heuristic decided a ridge
bifurcation was detected if the distance and direction of the two valleys changed.
Amengual, Pereze, Prat, Saez and Vilar (1997) described a three phase
approach for minutiae extraction. The first phase involved the processes of
segmentation, directional image computation, directional image smoothing and
binarization of image. The second phase involved thinning of the ridge lines in
order to reduce the number of undesirable structures such as spurs and holes in
the ridge lines. The minutiae extraction algorithm was based on method
proposed by Rafat and Xiao (1991). They also used the concept of a CN. The
CN for a pixel was computed using (Rafat et. al, 1991):
k=8
CN = ∑ | Γ(k + 1) −Γ(k) |
k =1
Γ(p) = 1
Γ(p) = 0
Eq. 2.2
where Γ (9)= Γ (1)
if pixel p is foreground
43
They used the following heuristic to classify the type of minutiae point:
if CN = 2 categorize as ridge ending
else if CN = 6 categorize as ridge bifurcation
The type of minutiae point, the x-coordinate, the y-coordinate, and the angle of
orientation was also extracted for each minutiae point.
Engler, Frank and Leung (1990) investigated a neural network classification
approach for extracting minutiae from noisy, gray scale fingerprint images. In this
approach gray scale fingerprint images were first passed through a rank of six
Gabor filters where each filter corresponded to a specific orientation of the
minutiae to be extracted. The outputs from the filters were fed into a 3 layered
back propagation neural network where the final output of the network was the
determination of presence of minutiae.
The National Institute of Standards and Technology (NIST) designed a minutiae
detection algorithm in the NIST Biometric Image Software (NBIS). Their minutiae
detection algorithm, called MINDTCT, was designed to operate on a binary
image where the black pixels represented ridges and white pixels represent
valleys on surface of the finger. A pixel was assigned a binary value based on
detection of ridge flow associated with the block. Detection of ridge flow in the
block indicated the need to analyze the pixel intensities surrounding the pixel
being examined, and a comparison with the grayscale intensities in the
surrounding region indicated if the pixel was part of a ridge or valley. Once the
binarization of the fingerprint image was complete, the binarized image was
scanned to identify localized pixel patterns. Candidate ridge endings were
detected in the binary image by scanning consecutive pairs of pixels in the
fingerprint image looking for sequences which matched the pattern indicated in
Figure 2.13.
44
*
Pattern
2X3
2X4
2X5
Figure 2.13 Pixel pattern used to detect ridge endings (Garris et al., 2004)
*
*
*
*
*
Ridge Ending
Ridge Ending
Bifurcation
Bifurcation
Bifurcation
appearing
disappearing
disappearing
appearing
disappearing
*
*
*
*
*
Bifurcation
Bifurcation
Bifurcation
Bifurcation
Bifurcation
disappearing
appearing
appearing
disappearing
appearing
Figure 2.14 Pixel patterns for minutiae detection (Garris et al., 2004)
Once a candidate minutiae point was detected, the type of minutiae point was
determined by comparing the pixel pair patterns as shown in Figure 2.14. All
45
pixel pair sequences matching these patterns formed a list of candidate minutiae
points.
The algorithm then performed post processing on the list of candidate minutiae
points to ascertain if they were spurious minutiae points and removed any
unnecessary artifacts in ridge lines, like islands and lakes (Figure 2.15).
Lake
Island
Figure 2.15 Unnecessary features (Garris et al., 2004)
2.4.5. Fingerprint Matching
Fingerprint matching is the process of finding a degree of similarity between two
fingerprints. There are different techniques of performing fingerprint matching,
each of which examines different features of the fingerprint. Currently there are
two primary methods used for fingerprint matching. The first method is a feature
based approach that uses minutiae details for matching purposes. The second
approach uses global ridge patterns of fingerprints for matching purposes. These
matching algorithms will output a match score, i.e. the likelihood of a given
fingerprint being the same as the one it is compared to. Fingerprint matching is a
challenging process because of inconsistencies introduced during interaction of
an individual with the sensor, temporary injuries to the surface of the finger,
inconsistent contact, and a variety of other issues. The following section
discusses research conducted previously in the field of minutiae based matching.
46
2.4.5.1. Experiments Related to Minutiae Based Matching
Chen, Jain, Karu and Ratha (1996) described a minutiae based matching
algorithm based on Hough Transform algorithm. The algorithm first estimated the
transformation parameters dx, dy, θ, and s between two fingerprint
representations. dx and dy were translations along the x and y-axis, θ as the
rotation angle, and s was the scaling factor. The second step involved aligning
the two representations and counting the number of matched pairs within a
margin of error. The matching scores for multiple transformations were collected
in an array, and the transformation which maximized the number of matched
pairs was considered to be the most accurate transformation. The final output of
the matching algorithm was a set of top 10 fingerprints that matched the input
fingerprint. They used a random sample of 100 fingerprints from the NIST-9
database to test the algorithm, and they achieved 80% accuracy rate at FRR of
10%. They also reported that their matching algorithm took about one second for
each match, and would require parallel processing for large scale matching
operations.
Bolle et al. (1997) described a string distance minutiae matching algorithm which
used polar coordinates of extracted minutiae features instead of their Cartesian
coordinates. This reduced the 2-D features to 1-D string by concatenating points
in an increasing order of radial angle in polar coordinates. Their algorithm
associated a ridge segment with each minutia which was used for aligning the
two fingerprints being compared. The rotation and translation which resulted in
maximum number of matched minutiae pairs as considered the most accurate
rotation. All minutiae points for the two fingerprints were converted into a 1-D
string at the optimal rotation angle, and the distance between the two strings was
computed. Tests of their algorithm on the NIST-9 database resulted in a FAR of
0.073% and FRR of 12.4% at threshold of seven, and FAR of0 .003% and FRR
of 19.5% at threshold level of 10.
47
A fingerprint matching algorithm is also included in NBIS. The matching algorithm
is a modified a version of a fingerprint matcher developed by Allan S. Bozorth
while at FBI. The matcher called BOZORTH3, uses only the x-coordinate, ycoordinate, and the orientation of the minutiae point to match the fingerprints.
The main advantage of this matcher is that it is rotation and translation invariant.
The main steps of the algorithm are as follows:
1. Minutiae features are extracted and an intra-fingerprint minutiae
comparison table is created for the input fingerprint and the
fingerprint to be matched against.
2. Use the minutiae table for the input fingerprint to the minutiae table
for the comparison fingerprint to create a new inter fingerprint
minutiae compatibility table.
3. Traverse the inter fingerprint minutiae compatibility table and
calculate a matching score.
Tests performed by NIST using the BOZORTH3 matcher showed a matching
score of 40 or higher as an acceptable threshold to consider the two fingerprints
to be from the same finger. Elliott & Modi (2006) performed a study using the
BOZORTH3 matcher which resulted in 1% FNMR for fingerprint dataset
comprised of individuals 18-25 years old.
2.5. Large Scale Systems and Distributed Authentication Architectures
Fingerprint systems vary in size depending on its application, and they are
typically characterized by the number of individuals in its database. Large scale
systems are characterized by high demand storage and performance
requirements, and currently large scale fingerprint systems retain fingerprint
templates for approximately 40 million people and process tens of thousands of
search requests everyday (Khanna, 2004).
The Registered Traveler (RT) program was authorized under the Aviation and
Transportation Security Act (ATSA) as a means to establish requirements to
48
expedite security screening of passengers (TSA, 2006). The RT program
biometric system architecture consists of a Central Information Management
System (CIMS), Enrollment Provider (EP), Verification Provider (VP), and
Verification Station. The EP collects biographic and biometric information from
applicants and transmits it to the CIMS. The CIMS aggregates the information
and returns the biometric and biographic information packaged so that it can be
stored on a smart card for the applicant. The EP has to use fingerprint sensors
approved by the FBI, but the VP is not constrained in their choice of fingerprint
sensors creating the possibility of enrolling and verifying on different fingerprint
sensors. Currently there are 5 airports participating in the pilot program with a
potential of more airports joining the program. Although the verification process
requires matching only against the template on the smart card, the issue of
interoperability still exists.
Large scale biometrics systems have to confront two different types of
challenges: those common to large scale information management systems, and
those specific to biometric implementations (Jarosz, Fondeur, & Dupre, 2005).
The main considerations of a large scale system can be categorized into the
following: architectural issues, administration issues, security issues, operational
issues, and performance issues. For any biometric system transaction
throughput and expected performance are of utmost importance, and the
acquisition subsystem, feature extraction and template generation subsystem,
storage subsystem, and matching subsystem contribute to those two factors. The
acquisition subsystem is the first point of interaction between a individual and the
system, and any inconsistencies introduced here will prorogate throughout the
system. For a large scale system based on distributed architecture, the issues
which arise at the first point of interaction become even more significant. The
following section describes the experiments performed related to fingerprint
sensor interoperability.
49
2.5.1. Fingerprint System Interoperability Experiments
Ko and Krishnan (2004) presented a methodology for measuring and monitoring
quality of fingerprint database and fingerprint match performance of the
Department of Homeland Security’s Biometric Identification System. They
proposed examining fingerprint image quality not only as a predictor of matcher
performance but also at the different stages in the fingerprint recognition system
which included: image quality by application, image quality by site/terminal,
image quality by capture device, image quality by new participants or repeat
visits, image quality by matcher enrollment, image quality by finger, and image
quality trend analysis. They also pointed out the importance of understanding the
impact on performance if fingerprints captured by a new fingerprint sensor were
integrated into an existing identification application. Their observations and
recommendations were primarily to facilitate maintenance and matcher accuracy
of large scale applications.
Marcialis and Roli (2004) described a study which utilized a multi-sensor system
for fingerprint verification. Along with examining interoperability error rates for
fingerprint datasets collected from optical and capacitive sensors, they proposed
a score fusion rule to decrease error rates for interoperable datasets. They used
Biometrika FX2000 optical fingerprint sensor and Precise Biometrics MC100
capacitive sensor. Both sensors provided 500 S4i resolution fingerprint images.
Using the FVC data collection process 10 fingerprint images were collected from
6 fingers of each of the 20 participants, resulting in a total of 1200 fingerprint
images on each sensor. The evaluation process consisted of four stages. In the
first stage the participant enrolled their index finger, middle finger, and ring finger
of both hands on a capacitive and optical fingerprint sensor using the same
minutiae extraction and template creation algorithm. In the second stage the
participants presented the same finger on both the sensors. In the third stage the
fingerprint acquired from the optical sensor in Stage 2 was matched to the
template created in Stage 1 from the optical sensor and the capacitive sensor.
50
This resulted in two sets of match scores for each participant. In the fourth stage
these two sets of match scores were fused to make a decision based on an
acceptance threshold. Their results showed a False Accept Rate (FAR) and
False Reject Rate (FRR) of 3.2% and 3.6% respectively for fingerprints from the
optical sensor, a FAR and FRR of 18.2% and 18.8% respectively for fingerprints
from the capacitive sensor, and a FAR and FRR of 0.7% and 1.3% respectively
for the fusion rule.
Jain & Ross (2004) investigated the problem related to fingerprint sensor
interoperability, and defined sensor interoperability as the ability of a biometric
system to adapt raw data obtained from different sensors. They defined the
problem of sensor interoperability as the variability introduced in the feature set
by different sensors. They collected fingerprint images on an optical sensor
manufactured by Digital Biometrics and a solid state capacitive sensor
manufactured by Veridicom. The fingerprint images were collected on both
sensors from 160 different individuals who provided four impressions each for
right index, right middle, left index, and left middle finger. They used a minutiae
based fingerprint matcher developed by Hong et al. (1997) to match images
collected from the sensor, and then compared images collected from Digital
Biometrics sensor to images collected from Veridicom sensor. Their results
showed that EER of 6.14%for matching images collected from Digital Biometrics
sensor and EER of 10.39% for matching images collected from Veridicom sensor
.The EER for the matching images collected from Digital Biometrics sensor to
Veridicom sensor was 23.13%. Their results demonstrated the impact of sensor
interoperability on matching performance of a fingerprint system. Nagdir & Ross
(2006) proposed a non-linear calibration scheme based on thin plate splines to
facilitate sensor interoperability for fingerprints. Their calibration model was
designed to be applied to the minutiae dataset and to the fingerprint image itself.
They used the same fingerprint dataset used in the study conducted by Jain &
Ross (2004), but used the VeriFinger minutiae based matcher and BOZORTH3
51
minutiae based matcher for the matching fingerprints. They applied the minutiae
and image calibration schemes to fingerprints collected from Digital Biometrics
sensor and Veridicom sensor and matched the calibrated images from the two
sensors against each other. Their results showed an increase in Genuine Accept
Rate (GAR) from approximately 30% to 70% for the VeriFinger matcher after
applying the minutiae calibration model. For the BOZORTH3 matcher an
increase in GAR from approximately 35% to 65% was observed.
Han et al (2006) examined the resolution different and image distortion due to
differences in sensor technologies. They proposed a compensation algorithm
which worked on raw images and templates. For raw images, all pixels of the
image were offset along horizontal upper, horizontal center, horizontal down,
vertical left, vertical center, and vertical right directions using the original
resolution of the fingerprint sensor. This compensation lead to a transformation of
location and angle of all minutiae points, along with a transformation of all pixels
in the image. For templates transformation, only the positions and angles of the
minutiae are transformed using the Unit Vector Method. Statistical analysis of
their experimental results showed a reduction in the mean differences between
non-compensated images and compensated images.
The International Labor Organization (ILO) conducted a biometric evaluation in
an attempt to understand the causes of the lack of interoperability (Campbell &
Madden, 2006). 10 products were tested, where each product consisted of a
sensor paired with an algorithm capable of enrollment and verification. Data was
collected from 184 participants and a total of 67,902 fingerprint images were
collected on all 10 products. The test was designed to be a multi visit study,
where the participant enrolled their left and right index finger and verified each
finger 6 times on each product during each visit. Native and interoperable FAR
and FRR were computed. Mean genuine FRR of 0.92% was observed at genuine
FAR of 1%. The objectives of this test were twofold: to test conformance of the
products in producing biometric information records in conformance to ILO SID-
52
0002, and to test which products could interoperate at levels of 1% FAR and 1%
FRR. The results showed that out of the 10 products, only two products were
able to interoperate at the mandated levels. The experimental design ensured
extraneous factors like temperature and ordering of products would have a
minimal effect on the results and outlined in detail in the report. This study was a
pure assessment test, and no attempt was made to analyze the lack of
interoperability.
The National Institute of Standards and Technology (NIST) performed an
evaluation test to assess the feasibility of interoperability of INCITS 378
templates. The objectives of the test were three fold:
1. To determine the level of interoperability in matching two INCITS 378
templates generated by different vendors.
2. To estimate the verification accuracy of INCITS 378 templates compared
to proprietary templates.
3. To compare the performance between the base INCTIS 378 template and
enhanced INCITS template which includes ridge count information.
The minutiae template interoperability test was called the MINEX 2004 test. Four
different datasets were used which were referred to as the DOS, DHS, POE and
POEBVA. The POE and POEBVA dataset were culled from the US-VISIT
dataset. All datasets had left and right index fingers using live-scan impressions.
All the datasets were operational datasets gathered in on-going US government
deployments. All the datasets combined consisted of fingerprint images from a
quarter of a million people were used, and approximately 4.4 billion comparisons
were executed in order to generate the performance evaluation reports. The
POEBVA dataset was collected in separate locations, in different environments
and different sensors. Also the data capture process for the authentication
images was unsupervised. Fourteen different fingerprint vendors participated in
the MINEX test.
53
The testing protocol examined three different interoperable verification scenarios:
1. The enrollment and verification templates were created using the same
vendor’s generator and template matcher from a different vendor.
2. The enrollment template generator and matcher were from the same
vendor and the verification template generator was from a different
vendor.
3. The enrollment template generator, the verification template generator and
template matcher were all from different vendors.
The performance evaluation criteria was based on false non match rates (FNMR)
and false match rates (FMR). FNMR was calculated at fixed FMR of 0.01% for all
the matchers. Performance matrices were created which represented FNMR of
native and interoperable datasets and provided a means for a quick comparison.
The tests did not show a particular pattern of better performance for comparison
of templates from the same generators compared to templates from different
generators.
The MINEX report identified quality of the datasets as a factor which affected
level of interoperability. The DOS and DHS datasets were of lower quality and
did not exhibit a level of interoperability as that of POEBVA and POE databases.
The MINEX test acknowledged it did not use images from sensors of different
technologies like capacitive and thermal. The test team also indicated that
performance for datasets which used fingerprints from capacitive, thermal and
optical sensors would be different than what was observed in the MINEX test.
One of the main conclusions from the MINEX test was related to the quality of
the datasets used in the test. Along with the quality metric, the environment
conditions and the interaction characteristics of data collection was not constant
for the fingerprint collection process. Environmental conditions and interaction
factors like sensor placement have an impact on the quality of fingerprints (Elliott,
Kukula, & Modi, 2007). The unknown nature of the data collection process was
not a part of the evaluation of the MINEX test.
54
Bazin & Mansfield (2007) described a study which focused on ISO/IEC 19794-2
standard for finger minutiae data exchange. The tests described in the paper
encompassed the first phase of Minutiae Template Interoperability Testing
(MTIT) evaluation with the aim of identifying key issues which contribute to
increased error rates in interoperability scenarios. This study analyzed compact
card templates along with data interchange record formats. The main difference
between record and compact card templates is based on representation of ridge
endings either by valley skeleton bifurcation points or by ridge skeleton end
points. In record templates ridge endings can be represented only by valley
skeleton bifurcation points. Four different vendors participated in this study and
the three interoperability verification scenarios used in MINEX test were used for
this study. Along with testing interoperability between vendors, this test also
examined level of interoperability between proprietary, record and card formats.
The two index fingers from 4,041 individuals were used in the final performance
evaluation test. As with the MINEX test, only optical live-scan sensors were used.
Also, fingerprints were also collected on 10-print scanners, and then segmented
for use in the test. In order to evaluate performance, FNMR performance
matrices were generated as several different fixed FMR. These matrices
contained native and interoperable dataset performance rates. The results from
this study showed lower FNMR for native matching comparisons compared to
interoperable matching comparisons at all levels of FMR.
As with the MINEX test, this test also examined the link between interoperability
and amount of data included in the standardized templates. The source of
fingerprints, characteristics of the fingerprint sensors, or characteristics of the
finger skin was not part of their analysis. The fingerprints were collected from
different environments which could potentially lead to extraneous variables
confounding results of the test. Not considering the source of the fingerprints
should not be seen as a failing of this test since it was not a part of their research
55
agenda, but the results point to the importance of studying fingerprint sensor
related issues as a means of reducing interoperability error rates.
Poh, Kettler and Bourlai (2007) attempted to solve the problem of acquisition
mismatch using class conditional score distributions specific to biometric devices.
Their problem formulation depended on class conditional score distributions, but
these probabilities were not known a-priori. They devised an approach which
used observed quality measures for estimating probabilities of matching using
specific devices. Their device specific score normalization procedure reduced the
probability of mismatch for samples collected from different devices. They
applied their approach to fingerprint and face recognition systems. They used the
BioSecure database for their study. They collected data from 333 participants, of
which 206 participants were considered to be genuine users. The remaining 126
participants were treated as zero-effort imposters. The fingerprint data was
collected on an optical fingerprint sensor and a thermal fingerprint sensor. The
BOZORTH3 matcher was used. They used the fingerprint quality indices
described by Jain, Chen, & Dass (2005). Using Expected Performance Curves
(EPC), which provided unbiased estimates of performance at various operating
points, their results showed better performance for matching fingerprints from
different devices using their framework.
2.6. Performance Metrics
Error rates are the basic units of performance assessment for matching engines.
These error rates not only reflect the strength of the matching algorithm, but also
the probabilistic nature of the decision making part of the algorithm. Attempting to
compare performance metrics for multiple matching engines has been a
challenge that researchers have attempted to address. The following sections
outline different performance metrics that map functional and operational
relationships between different error rates.
56
2.6.1. Receiver Operating Characteristic Curves
Receiver operating characteristic (ROC) curves are a means of representing
results of performance of diagnostic, detection and pattern matching systems
(Mansfield & Wayman, 2002). A ROC curve plots as a function of decision
threshold the FMR on the x-axis and true positive rates on the y-axis.
ROC (T) = (FMR (T),TAR (T)) where T is the threshold
Eq. 2.3
For performance comparison of biometric systems, a modified ROC curve called
Detection Error Tradeoff (DET) curve is used (Doddington, Kamm, Martin,
Ordowski, & Przybocki, 1997). A DET curve will plot FAR on the x-axis and FRR
on the y-axis as function of decision threshold. A DET curve can also be created
by plotting FMR on the x-axis and FNMR on the y-axis as function of decision
threshold. Comparing DET curves for different systems allows comparison of the
systems at a threshold that is deemed preferable for the application of the
biometric system.
DET (T) = (FMR (T), FNMR (T)) where T is the threshold
Eq. 2.4
2.6.2. Equal Error Rate
Using multiple DET curves to compare matching performances of different
matchers assumes having a known operational threshold. A more convenient
comparison of multiple matchers would necessitate the reduction of the DET
curve to a single number (R. M. Bolle, Connell, Pankanti, Ratha, & Senior,
2004a). The EER operating point is a computation which is generally regarded as
an obvious choice to judge quality of a matcher. The EER is the operational point
where FNMR=FMR. This is a useful point of comparison only if the FNMR and
FMR are supposed to be equal. For unequal FNMR and FMR, analysis at EER
operational point will not provide any useful information.
57
2.6.3. Difference in Match and Non-Match Scores
Measuring the difference in match score density and non-match score density
has been used as a performance assessment criteria. Daugman and Williams
(1996) define the measure of separation for a matcher as
d' =
μmatchscores − μnonmatchscores
Eq. 2.5
(1/2) (σ 2 matchscores + σ 2 nonmatchscores)
µmatchscore and σ2matchscore is mean and variance of match scores of genuine users,
and µnon-matchscore and σ2non-matchscore is mean and variance of non-match scores of
mismatch fingerprints, respectively. d’ can be used to compare multiple
matchers, but it will be reliable only if there is a significant difference in
performance. Matchers that might have similar genuine score distributions can
be hard to compare, but using d’ measure is useful because it uses non-match
score distributions as part of the computation.
2.6.4. User Probabilities and Cost Functions
The previous two methodologies assume that the probability of FMR and FNMR
is the same. A particular biometric system might have to be analyzed keeping in
mind the probability of a user being an imposter or a genuine user, and the
consequences of a false match or a false non-match. There are methodologies
which use security and convenience tradeoff to determine which system is better
suited for particular requirements.
In Equation 2.6 PG is the prior probability of a user being genuine, CFNM is the
cost associated with a false non match, CFM is the cost associated with a false
match (R. M. Bolle, Connell, Pankanti, Ratha, & Senior, 2004a). This is a useful
method for quantifying the cost of the system in dollars, but it is only as accurate
as the input parameters.
58
Cost = CFNM * FNMR * PG + CFM * FMR * (1-PG)
Eq. 2.6
2.6.5. Cumulative Match Curve
Cumulative Match Curve (S3C) is a statistic that measures capabilities of
matching engine that returns a ranked list. The ranked list can be interpreted as a
list of enrolled samples that are arranged according to the match scores against
an input sample. A better matching engine will return a better ranked list. ROC
curves and its derivatives are useful for analysis of matching engines that
produce continuous random variables; a S3C is useful for analysis of matching
engines that produce discrete random variables (R. M. Bolle, Connell, Pankanti,
Ratha, & Senior, 2004b)
59
CHAPTER 3. METHODOLOGY
3.1. Introduction
The chapter outlines the data collection, data processing, and data analysis
methodology utilized in this dissertation. The main purpose of this chapter is to
ensure the experiment can be repeated in a reliable manner.
3.1.1. Research Design
This dissertation used the experimental research method. Experimental research
methods require a highly controlled environment with a clear identification of
independent variable/variables, dependent variables, and extraneous variables.
In this study the fingerprint sensors were the independent variables and minutiae
count, image quality scores, and match scores were the dependent variables.
This study also required an experimental research method to ensure all subjects
followed the same process of providing fingerprints on different sensors. The
experimental research method is useful for controlling sources of error. For this
experiment, the sources of error were controlled by randomizing assignment of
sensors to subjects, maintaining repeatability of the process, and controlling the
extraneous variables like room lighting, temperature, and humidity. The
experimental research method also allowed for an experimental design which
minimized internal and external validity issues. These are discussed in Section
3.5.
60
3.2. Data Collection Methodology
Each participant was required to follow the same data collection procedure. At
the beginning of the experiment, the participant completed a questionnaire, the
details of which are listed in Section 3.2.8. The responses were collected to
provide further research opportunities of investigating their correlation with image
quality and performance metrics. The detailed questionnaire can be found in
Appendix B.
After completing the questionnaire, the participant proceeded with the fingerprint
data collection procedure. A test administrator supervised the data collection.
There were nine fingerprint sensors used in this study, and their details can be
found in Section 3.2.4. The participant interacted with each fingerprint sensor
while sitting in a chair. The participant provided six fingerprint images from the
index finger of his/her dominant hand for each sensor. Previous studies have
showed that index fingers tend to provide higher quality images compared to
other fingers (Elliott & Young, 2007). If a participant was ambidextrous, the index
finger from the preferred writing hand was used. The participant was given a
randomized list of fingerprint sensors and started the data collection with the first
sensor on the randomized list. Before interacting with each fingerprint sensor, the
participant’s moisture content, elasticity, oiliness and temperature was recorded
from the surface of the finger skin. The fingerprint sensor was placed on a
pressure measuring device. The peak pressure applied by the participant’s finger
on the sensor while providing each fingerprint image was recorded. A detailed
description of the hardware used in this study is given in Sections 3.2.3 and
3.2.4. Data collection was complete only after the participant provided
fingerprints on all sensors. A flowchart of the data collection protocol is given in
Appendix A. The physical layout of the data collection area is given in Appendix
D.
61
3.2.1. Participant Selection
Participants were selected from the population available at the West Lafayette
campus of Purdue University. For the purpose of this study convenience
sampling was performed to form the participant pool. 190 participants completed
the study.
3.2.2. Timeline
Data collection for this study commenced on January 31st, 2008 and was
completed on April 25th, 2008.
3.2.3. Data Collection Hardware & Software
Moisture content, oiliness, and elasticity finger skin were measured using
Triplesense TR-3 manufactured by Moritex Corporation (Figure 3.1). This is a
commercially available hand held device capable of providing all three
measurements. Measurements are on a normalized scale of 1-100 based on
testing performed by the manufacturer. This device came pre-calibrated from the
manufacturer. No additional calibration was performed on the device.
The temperature of finger skin was measured using Raytek MiniTempTM Infrared
Thermometer (Figure 3.2). This is a commercially available device which
provides readings on a Fahrenheit scale. This device came pre-calibrated from
the manufacturer and no additional calibration was performed.
62
Figure 3.1 Triplesense TR-3
Figure 3.2 Raytek MiniTempTM
A Vernier Force Plate was used to measure the force applied on the fingerprint
sensor. The device was calibrated to zero N before each interaction.
63
3.2.4. Fingerprint Sensors
Fingerprints were collected using nine different sensors. There were three swipe
and six touch interaction type sensors. There was one thermal sensor, four
capacitive sensors and four optical sensors. All sensors were of approximately
500 S4i resolution. All fingerprint sensors chosen were made by commercial
sensor manufacturers and were commercially available. None of the fingerprint
sensors were specifically built for this dissertation. The fingerprint sensors were
also chosen on basis of availability and with the intent of including sensors that
were of swipe and touch interaction type and optical, thermal, and capacitive
acquisition technology. There are other types of acquisition technologies
available which were not selected due to their limited use in the live deployments.
The specifications for each sensor are listed in Table 3.1.
Table 3.1 List of Fingerprint Sensors
Fingerprint
Sensor
Sensor 1
Sensor 2
Sensor 3
Sensor 4
Sensor 5
Sensor 6
Sensor 7
Sensor 8
Sensor 9
Type of
Sensor
Action
Capture
Area
(mm)
14 X .4
13.8 X 5
30.5 X 30.5
14.6 X 18.1
12.8X15
16 X 24
15 X 15
12.8 X 18
12.4 X .2
Resolution
S4i
(approximate)
500
500
500
500
500
500
500
508
500
3.2.5. Fingerprint Sensor Maintenance
The fingerprint sensors were calibrated the first time they were setup for the
experiment. Calibration was performed using the software which accompanied
each sensor. The fingerprint sensors were not recalibrated after the first setup.
64
At the beginning of each fingerprint collection session, every fingerprint sensor
was cleaned by wiping it with a piece of cloth (Campbell & Madden, 2006). This
action was performed to minimize any accumulation of residue.
3.2.6. Software or Sensor Malfunction
In case of hardware or software malfunctions, the following steps were taken:
•
Restart the data collection software.
•
Reconnect and reinitialize the fingerprint sensor.
•
Reboot the data collection computer.
In case of a sensor failure, all fingerprints collected from that sensor were to be
removed from the data analysis section. None of the fingerprint sensors used for
data collection failed during the data collection period.
3.2.7. Variables Measured during Data Collection
Fingerprint images collected from all sensors were stored for the duration of the
study. Information about the participant’s age, gender, ethnicity, occupation,
missing fingers, temperature of finger surface, moisture content of finger surface,
oiliness of skin, elasticity of skin and pressure applied on the sensor was
collected as performed in a previous study by Kim et al. (2003). Failure to acquire
(FTA) was recorded if Verifinger 5.0 extractor determined a fingerprint image to
be of insufficient quality. If three consecutive failure to acquire attempts occurred,
a failure to enroll (FTE) was recorded and the participant was asked to proceed
with the next fingerprint sensor on the randomized list. For the purpose of this
study gender, ethnicity, and occupation were treated as categorical variables;
missing fingers, temperature, moisture content, oiliness and elasticity of finger
skin, and pressure applied on fingerprint sensor were treated as interval
variables; age was treated as a ratio variable. This information was collected to
report the demographic mix of the participants, but was not used for analysis.
65
3.2.8. Variables Controlled during Data Collection
Each participant had the discretion of adjusting the height of the chair during the
data collection session. The lighting level of the room, room temperature, relative
humidity in the room was kept within the Knoy Hall, West Lafayette, building
regulations during the experiment. The tilt angle of the fingerprint sensor was
controlled by laying it flat on the pressure plate surface. The surface of the
pressure plate and the back of each fingerprint sensor was covered with Velcro
to minimize the movement of the sensor when the participant interacted with the
sensor. The surfaces of all sensors were cleaned at the start of each data
collection session.
3.3. Data Processing Methodology
3.3.1. Minutiae Count and Image Quality Processing
The basic variables analyzed for this study were minutiae count, and image
quality score of the fingerprint image. Minutiae count and image quality scores
were generated using Aware Image Quality, and MINDTCT and NFIQ which are
a part of NBIS. Minutiae count was also generated using VeriFinger 5.0 extractor.
For Aware Image Quality Software, minutiae count and image quality scores
were ratio variables. Minutiae count was always greater than 0, and image
quality scores ranged from zero to 100. For MINDTCT, minutiae count was a
ratio variable and was always greater than zero. NFIQ scores ranged from one to
five and was treated as an interval variable. For VeriFinger 5.0 minutiae count
was a ratio variable. The determination of the variable types was used to
formulate the appropriate statistical tests.
66
3.3.2. Fingerprint Feature Extractor and Matcher
Fingerprint feature extractors and fingerprint matchers provided with VeriFinger
5.0 and NBIS were used. VeriFinger 5.0 is a commercially available fingerprint
feature extractor and matcher. It has been a participant in Fingerprint Verification
Competitions in 2000, 2002, 2004, and 2006, and in the INCITS 378 Fingerprint
Template performance and interoperability test conducted by NIST. MINDTCT
and BOZORTH3, which are a part of NBIS, were used for extraction and
matching respectively. The genuine match scores and imposter non-match
scores were also ratio variables and were always a non-negative number.
3.4. Data Analysis Methodology
Hybrid testing, which combined live acquisition and offline matching, was
performed to analyze the data collected in this experiment (Grother, 2006).
Genuine match scores and imposter match scores were generated offline once
all 190 participants had completed their data collection sessions. A hybrid testing
scenario was necessary for an experiment which incorporated multiple sensors
because live users would not want to sit through combinatorial use of multiple
sensors. The scenario also provided an opportunity to use the entire dataset for
testing purposes.
3.4.1. Score Generation Methodology
The process outlined in this section was performed using VeriFinger 5.0 and
BOZORTH3. All participants provided six fingerprint images on each sensor from
the index finger of their dominant hand, or their preferred writing hand if they
were ambidextrous. The resulting six fingerprint images participant were split into
two groups; the first three images were placed in an enrollment dataset and the
last three images were placed in a test dataset. Enrollment template datasets
were created for all nine sensors, and test template datasets were created for all
nine sensors using feature extractors from VeriFinger 5.0 and BOZORTH3. This
67
methodology is graphically represented in Figure 3.3. Enrollment templates from
each dataset were compared against templates from each test dataset resulting
in a set of scores S, where
S ={(Ei,Vj,scoreij)}
i= 1,.. ,number of enrolled templates
j = 1,.., number of test templates
scoreij = match score between enrollment template and test template
The set of scores, which consisted of genuine match scores and imposter match
scores, were placed in a matrix form (Figure 3.4). This matrix of scores had the
following properties:
1. The number of rows and columns were equal, and corresponded to the
number of fingerprint sensors used.
2. The diagonal represented the native datasets, and all the cells not in the
diagonal represented interoperable datasets.
68
Figure 3.3 Generation of match scores
69
Figure 3.4 Scores of native and interoperable datasets
3.4.2. Analysis Techniques
The steps outlined in this section were performed on data generated from Aware
Image Quality Software, VeriFinger 5.0, MINDTCT and BOZORTH3. Based on
previous research conducted in fingerprint sensor interoperability experiments,
the analysis methodology was categorized into the following:
1. Basic fingerprint feature analysis.
2. Match scores analysis.
3.4.2.1. Basic fingerprint feature analysis
An important factor when considering interoperability is the ability of different
sensors to capture similar fingerprint features from the same fingerprint. Although
human interaction with the sensor introduces its own source of variability, it is
nonetheless important to statistically analyze variability introduced from different
sensors. The flowchart of the fingerprint feature statistical analysis methodology
performed in this dissertation is shown in Appendix C.
70
The analysis described in this section was performed using all six fingerprints
collected from the participant. The basic fingerprint feature analysis involved
examination of minutiae count and image quality scores. Minutiae count and
image quality scores were calculated for each fingerprint image using Aware
Image Quality, MINDTCT and NFIQ. Statistical tests were performed to test
similarities in minutiae count between all fingerprint datasets. This analysis was
performed separately on minutiae count generated by Aware Image Quality and
MINDTCT. Statistical tests were performed to test similarities in image quality
between all fingerprint datasets. This analysis was performed separately on
image quality scores generated by Aware Image Quality and NFIQ. A model
adequacy check for a parametric F-test was performed which involved the
following:
1. Normality of residuals.
2. Constancy of variance of error terms.
3. Independence of observations.
For a parametric F-test of a between groups effect, it is more appropriate to
check normality of the residuals than the raw scores (Montgomery, 1997). A
parametric F-test is a ratio test of between group variance and within group
variance. The within groups variance is the average square of the residuals of
the group mean. A test of normality of residuals will show if any particular
observation has a very strong influence on the group mean compared to other
observations. This ensures that scores which are used to calculate the withingroup variance are normally distributed. A violation of any of the assumptions 1)
or 2) can be fixed by transformation of data (Montgomery, 1997).
Two sets of hypotheses were formed to test for similarity of minutiae count and
image quality. The first set of hypothesis was used for testing similarity of
minutiae count.
71
H10: There is no statistically significant difference among the mean minutiae
counts of all fingerprint datasets.
H1A: There is a statistically significant difference among the mean minutiae
counts of any fingerprint datasets.
H10: µi minutiae_count = µ2 minutiae_count……..= µn minutiae_count
Eq. 3.1
H1A: µi minutiae_count ≠ µ2 minutiae_count……..≠ µn minutiae_count
n = number of datasets
The second set of hypothesis was used for testing similarity of image quality
scores.
H20: There is no statistically significant difference among the mean image quality
scores of all fingerprint datasets.
H2A: There is a statistically significant difference among the mean image quality
scores of any of the fingerprint datasets.
H20: µi qscore = µ2 qscore……= µn qscore
Eq. 3.2
H2 A: µi qscore ≠ µ2 qscore…..≠ µn qscore
If a statistically significant difference was observed for the test, all possible pairs
of means were compared. Tukey’s Honestly Significant Difference (HSD) was
used to test all pairwise mean comparisons. Tukey’s HSD is effective at
controlling the overall error rate at significance level α, and thus preferred over
other pairwise comparison methods (Montgomery, 1997). Two sets of
hypotheses were tested for pair-wise differences in means, one set of hypothesis
corresponding to the similarity of minutiae counts, and one set corresponding to
the similarity of image quality scores.
H30: µi minutiae_count = µj minutiae_count for all i ≠ j
H3 A: µi minutiae_count ≠ µj minutiae_count for all i ≠ j
Eq. 3.3
72
H40: µi qscore = µj qscore for all i ≠ j
Eq. 3.4
H4 A: µi qscore ≠ µj qscore for all i ≠ j
3.4.2.2. Match Scores Analysis
The following analysis was performed separately for genuine match scores and
imposter match scores generated by VeriFinger 5.0 and BOZORTH3. The set of
genuine match scores and imposter match scores were analyzed with the data
from the matrix of set of scores shown in Figure 3.5. The performance of
interoperable datasets was analyzed using the following methods:
1. Performance interoperability matrix consisted of FNMR for all datasets
collected in the experiment (Campbell & Madden, 2006). The VeriFinger
5.0 matcher was used to generate the set of raw scores and FNMR at
fixed FMR of 0.01% and 0.1%. The VeriFinger 5.0 matcher deduces the
decision threshold of FMR operational points based on internal tests of the
matcher. For each fixed FMR a FNMR interoperability matrix was
generated (Figure 3.5). These two matrices had exactly the same
properties as the matrix of set of scores, as outlined in Section 3.4.1. For
the BOZORTH3 matcher a single decision point was used to create the
interoperability performance matrix. A match score of 40 or above
indicated a true match as suggested by Garris et al. (2004). Since the
score of 40 does not correspond to a particular FMR operational point, no
operational point was assigned to performance interoperability matrix
generate by NBIS matcher scores.
73
Figure 3.5 Performance interoperability matrix at different operational FMR
For each interoperable dataset a simple placement consistency metric
was formulated. If VeriFinger 5.0 extractor detected a core in the pair of
fingerprints being matched by VeriFinger 5.0 matcher, the placement was
considered to be consistent. Using this measure, a percentage of
fingerprint matching pairs in which both had a core was calculated for all
interoperable datasets. This analysis was performed to analyze the
relation between consistency of placement and interoperable FNMR.
74
2. Test of FNMR homogeneity was performed on FNMR calculated in the
previous section to test for equality using the chi-square distribution
testing for homogeneity of proportions. The main objective of this test was
to examine if the difference in FNMR among the datasets was statistically
significant in their difference. This test aided in statistically testing equality
of FNMR for native datasets and interoperable datasets. Since there were
nine sensors, there were nine corresponding sets of hypothesis. Eq 3.5
shows a sample of the null hypothesis and alternate hypotheses for
dataset S1:
H10: pi = p2 = ……..= pn
Eq. 3.5
H1A: pi ≠ p2 ≠ ……..≠ pn
n = total number of native and interoperable datasets for sensor 1.
p= FNMR.
The chi-square test statistic is calculated as follows:
( fo − fc) 2
2
χ =∑
fc
n
Eq. 3.6
fo = observed frequency
fc= theoretical frequency if no false non matches occurred
The critical value of χ2 was computed at a significance level of 0.05 and
degrees of freedom (n-1), and then compared to the test statistic χ2 from
Eq. 3.6. If the test statistic exceeded the critical value, the null hypothesis
was rejected. This same hypothesis test was repeated for the other eight
fingerprint sensors. If the null hypothesis was rejected, further analysis
was performed to examine which interoperable datasets caused the
rejection. This was done using the Marascuillo procedure for a
significance level of 0.05. The Marascuillo procedure simultaneously tests
75
the differences of all pairs of proportions for all groups under investigation.
The critical range was calculated as shown in Eq. 3.7 and compared to the
result of Eq. 3.8. If the result from Eq. 3.8 exceeded the critical range, the
difference was statistically significant.
rij = χ
2
(α , k −1)
pi (1 − pi ) p j (1 − p j )
+
ni
nj
Eq. 3.7
ni = sample size of group i
pi = FNMR of group i
k = total number of sensor groups
| pi − p j |
Eq. 3.8
3. Test of Match of Scores: In order to statistically analyze the differences in
genuine match scores and imposter match scores between the native
dataset and interoperability dataset, either the Dunnet’s comparison
method (Montgomery, 1997) or the Kruskal Wallis test was used
depending on model adequacy checks described in Section 3.4.2.1.
Dunnet’s method is a modified form of a t-test. The mean genuine match
score and the mean imposter match score for the native dataset would be
considered as the control. The mean genuine match score and mean
imposter match score for each interoperable dataset would be tested
against the control (i.e. the native dataset scores). One measure of
fingerprint sensor interoperability can be described as the consistency of
match scores of the matcher for native fingerprint datasets and
interoperable fingerprint datasets. In statistical terms, this would be
examined by testing for a significant difference of the genuine match
scores between native and interoperable fingerprint datasets. The same
76
test would be repeated to test for a significant difference of imposter
match scores between native and interoperable fingerprint datasets.
3.4.3. Impact of Image Quality on Interoperable Datasets
The impact of image quality on interoperable datasets was examined by
removing low quality images and recalculating the test of FNMR homogeneity.
Only fingerprint images which had NFIQ quality scores of one, two or three were
used for the matching operation. The NIST Minutiae Exchange Interoperability
Test report indicated that a relative reduction in interoperable dataset FNMR
would be observed since poorly performing images were not used (Grother et al.,
2006). This analysis was performed to examine if using higher quality images
would lead to a higher degree of similarity in FNMR between native and
interoperable datasets. After fingerprint images with NFIQ score of one, two or
three were selected from the interoperable datasets and VeriFinger 5.0 template
generator and matcher was used to perform the matching operations.
The test of homogeneity of proportions was conducted on the FNMR of native
and interoperable datasets comprised of fingerprint images with NFIQ score of 1,
2 or 3. This was repeated for all possible interoperable datasets.
3.4.4. Post Hoc Analysis
The moisture content, oiliness, elasticity, and temperature of the finger skin were
measured before an individual interacted with each new sensor. Basic
descriptive analysis was performed on these variables to analyze their range,
mean and variance. An exploratory analysis involving correlation matrices was
performed on moisture content, oiliness, elasticity and temperature of the finger
skin and its relation to image quality of the fingerprint provided. This analysis can
be found in Appendix J.
77
An exploratory analysis using correlation matrices was also performed between
all of the following relations: match scores of a pair of fingerprint images and the
difference in quality of the pair of images, difference in ridge bifurcation count of
the pair of images, difference in ridge ending count of the pair of images, and
different types of acquisition technologies that were used to capture the pair of
fingerprint images.
3.5. Threats to Internal and External Validity
3.5.1. Internal Validity
For the purpose of this study internal validity was defined as the confidence
placed in the cause and effect relationship between the independent variable and
dependent variables without the relationship being influenced by extraneous
variables (Sekran, 2003). The following seven internal validity issues are
discussed: history, selection, maturation, instrumentation, mortality rate, repeated
testing, and experimenter bias.
History
Outside events may influence participants during the experiment or between
repeated measures of the dependent variable (“Psychology 404,” 1998).
Biometrics is currently receiving a lot of media coverage because of privacy
issues. These events could have influenced a participant’s willingness to
participate or comply with the instructions. Although these events are beyond the
control of the experimenter, the participants were given an opportunity to ask
questions about the experiment before proceeding with it. History effects are a
more relevant threat to experiments which require repeated measures spread
over multiple sessions. Since this was a single session experiment history effects
were deemed to be minimal.
78
Selection
The threat of selection arises from selection of participants for multiple groups in
an experiment. Experiments which employ a control group and different test
groups are susceptible to this threat, as they utilize specific criteria for placing
participants in certain groups. This experiment did not have multiple groups of
participants -thus this threat was not a concern.
Maturation
For the purpose of this study, maturation was defined as the difference in
measurements of the dependent variable due to passage of time (“Psychology
404”, 1998). Participants becoming more skillful, observing new details of the
experiment, and acclimatizing themselves to the experimental conditions can
affect the variables of interest. Multiple visit studies or studies with a long gap
between subsequent visits are vulnerable to this threat. In this case participants
completed the study only once, which reduced the risk of maturation. Participants
waiting to complete the study could observe the previous participant, potentially
causing the participants to change their interactions with the fingerprint sensors.
For this reason, the data collection room was kept separate from the waiting
room, and this threat did not have an impact on data collection.
Instrumentation
The reliability of instruments used to capture, measure and gauge the variables
of interest is of utmost importance to an experimental setup. In order to have
confidence in the results, the instruments should provide consistent readings
throughout the study. All instruments used in this study are commercially
available products, and remained the same for all participants. The fingerprint
collection software, fingerprint feature extraction software, and fingerprint
matching software are all commercially available or open source. A PC with
Windows XPTM operating system was used for this study, which was not changed
for the duration of data collection.
79
Mortality Rate
For the purpose of this study, mortality rate was defined as the dropout rate of
participants from the experiment. Participants completed the study in a single
session. The participants did not have to return for repeated sessions. This
reduced the threat of mortality. Participants had the option of terminating the data
collection session at any time during the study. Although this posed a constant
threat to internal validity, the threat was mitigated by providing a positive and
unambiguous experience to the participant.
Repeated Testing
Repeated testing can confound the measurement of dependent variables as the
participants get repetitive in their experimental interaction and the variables of
interest are not affected by manipulation of independent variables (Sekran,
2003). The experimental setup of this study required each participant to
participate in a single session which reduced the threat of repeated testing.
During this session, the participants were asked to provide six fingerprint images
from the index finger of their natural hand on all fingerprint sensors. It was
infeasible to completely control for repeated testing in such a setup. In order to
minimize the order effects introduced by repeated interaction, each participant
was presented with a random ordering of fingerprint sensors. This reduced the
chances of random error affecting only one participant.
Experimenter bias
Expectations of an outcome by persons running the experiment could
significantly influence the outcome of the experiment (“Psychology 404”, 1998).
The data recording in this study was automated except for one fingerprint sensor
which required the administrator to click the capture button. The ability of the
administrator to influence the fingerprint placement of a participant was a
constant threat, but it existed equally for all participants thereby minimizing its
ability to affect specific participants. The generation of results for analysis was
80
automated by the fingerprint feature extractor and matcher thereby reducing the
threat of experimenter bias.
3.5.2. External Validity
For the purpose of this study, external validity was defined as the ability to
generalize the results from a study across the majority of the experimental
population (Sekaran, 2003). Threats to external validity can arise from population
difference, experimental setup, and localization issues. These issues and their
respective mitigation methods are discussed in detail in the following sections.
Population Difference
Population difference arises when samples selected for the experiment are not
representative of the population to which the results are generalized. This threat
can typically be mitigated by choosing a sample that is representative of the
population. Data collection for this study was conducted at the West Lafayette
campus of Purdue University. A majority of participants were students attending
Purdue University, and although participants from a diverse age group were
sought for this experiment in order to make the sample makeup more
representative of the population, this external validity issue constrained
generalizability of this study. Previous studies have shown that there is an age
effect on fingerprint recognition performance, and therefore the generalizability of
this study was restricted to a similar aged population (Elliott & Sickler, 2005).
Experimental Setup
Fingerprint recognition can be used for logical or physical access control. The
deployment environments for these two types of access control can be vastly
different. Physical access control systems are placed in a relatively uncontrolled
environment and the interaction postures also differ vastly from logical access
control. Logical access control systems are used with some kind of a computing
resource, like a desktop, and are generally in a controlled environment. In order
81
to control the environmental threat, the experiment was designed for logical
access control applications used for network or desktop login. Participants
interacted with the fingerprint devices while seated in a chair, and were allowed
to adjust the seat height as to their preference. This setup was representative of
a real world scenario for logical access control in an office environment. The
temperature and lighting in the experimental area were controlled so that their
effects were representative of an office environment.
Localization Issues
Applying the results from an experiment conducted in a laboratory to a general
operational scenario can lead to localization issues. This threat is more relevant
to experiments which can be affected by a change in geographical locations. This
experiment was conducted in a highly controlled environment which simulated an
office environment. Office environments across different geographical regions
tend to have similar conditions, so the results of this study are generarlizable
across office environments irrespective of their locations. The skin on fingers is
affected by outdoor environmental conditions, and this can affect fingerprint
recognition systems. For the purpose of this experiment this was not expected to
be severe due to the ability of the human body to acclimate itself to a controlled
environment. Localization effects were not anticipated to be a major threat for
this study.
3.6. Evaluation Classification
Table 3.2 partitions the test along seven categories and provides a summary of
the experimental evaluation.
82
Table 3.2 Evaluation Classification (Mansfield & Wayman, 2002).
Experimental Application Types
Application Classification
Co-operative or Non Co-operative
Overt versus Covert
Habituated versus Non-Habituated
Attended versus Non-Attended
Standard Environment
Public versus Private
Open versus Closed System
Classification for this research
Technology
Co-operative Users
Overt
Both
Attended
Yes
N/A
Closed
3.7. Summary
This chapter has outlined the data collection, data processing and data analysis
methodology that were used in this dissertation. Both the data collection
procedures and test methodology for this study were based on prior research
related to interoperability of fingerprint sensors and fingerprint recognition
systems. This chapter also acknowledged the different extraneous variables that
potentially affected the validity of the study, and explained how the study
attempted to mitigate the effects of the variables to increase validity of the study.
83
CHAPTER 4. DATA ANALYSIS
This chapter discusses results from surveys completed by the participants,
followed by analysis of minutiae count, image quality scores and match scores of
native and interoperable datasets. The data analysis methodologies and
statistical tests described in Chapter 3 were performed on the datasets collected
from the nine sensors. Each native dataset was named after the sensor that was
used to collect it. Table 4.1 outlines the details of each sensor and the dataset
name given to each sensor.
Table 4.1 Coding of Fingerprint Sensors and Datasets
Dataset Name
S1
S2
S3
S4
S5
S6
S7
S8
S9
Fingerprint
Sensor
Type of
Sensor
Action
Swipe
Swipe
Touch
Touch
Touch
Touch
Touch
Swipe
Touch
Capture
Area
(mm)
14 X .4
13.8 X 5
30.5 X 30.5
14.6 X 18.1
12.8X15
16 X 24
15 X 15
12.4 X .2
12.8 X 18
Throughout this chapter, interoperable datasets were named using the notation
{ES,TS} where ES refers to the source of the enrollment images and TS refers to
the source of the test images. For example, {S1, S2} refers to S1 Fingerchip
being the source of enrollment images and S2 being the source of test images.
84
Results of the survey filled out participants at beginning of the data collection
session can be found in Appendix I.
4.1. Failure to Enroll (FTE)
Data acquisition was performed using the native acquisition software provided by
the sensor manufacturers. All capture software except for S3 LC 300 used an
auto capture acquisition mode and had an inbuilt quality assessment algorithm
which made a determination of the acceptability of the fingerprint image. Failure
to Enroll (FTE) was not based on acceptability of the fingerprint image by the
native software but instead by VeriFinger 5.0. FTE was recorded for an individual
if three fingerprint images could not be processed using VeriFinger 5.0 extractor.
Determination of FTE was conducted in offline processing mode. S3 LC 300
could not to be used in auto capture acquisition mode. The determination of
capturing the fingerprint image using S3 LC 300 was made by the test
administrator based on subjective assessment of the clarity of the fingerprint
image. Table 4.2 shows the number of FTE for each sensor. Fingerprints
collected using the S5 sensor had the highest FTE out of all the sensors.
Table 4.2 Number of participants that recorded Failure to Enroll (FTE) using
VeriFinger 5.0
Sensor
Sensor 1
Sensor 2
Sensor 3
Sensor 4
Sensor 5
Sensor 6
Sensor 7
Sensor 8
Sensor 9
N
0
3
0
2
5
1
1
2
3
FTE Rate
0%
1.5%
0%
1.0%
2.6%
0.5%
0.5%
1.0%
1.5%
85
4.2. Basic Fingerprint Feature Analysis
This section describes results of minutiae count analysis and image quality
analysis performed on fingerprint images collected from all nine fingerprint
sensors. The dataset effects were treated as the main effects for all statistical
tests.
4.2.1. Minutiae Count Analysis
Three different software were used to extract the minutiae count: Aware Quality
Software, MINDTCT, and VeriFinger 5.0 extractor. The results of these three
different software were not compared to one another.
The preliminary test of assumptions for a single factor parametric F- test was
performed on minutiae count generated by Aware Quality Software. A model
adequacy check was performed by examining the residual values. A visual
analysis showed that the normal probability plot, independence of samples, and
constancy of variance did not violate the assumptions for performing the
parametric F-test. These graphs can be found in Appendix E. Table 4.4 shows
the values of mean (M) and standard deviation (SD) for each of the fingerprint
datasets. Figure 4.1 shows the boxplot of minutiae count for all datasets and
Figure 4.2 shows the histogram of minutiae count.
Table 4.3 Descriptive statistics for Aware software minutiae count analysis
Dataset
S1
S2
S3
S4
S5
S6
S7
S8
S9
n
1140
1122
1140
1128
1110
1134
1134
1128
1122
M
72.91
31.72
39.93
39.48
29.16
46.50
32.68
35.39
47.65
SD
18.16
11.61
13.67
10.57
8.92
10.83
9.16
10.39
11.57
86
Figure 4.1 Boxplot of Aware software minutiae count
Figure 4.2 Histogram of Aware software minutiae count
87
The omnibus test for main effect of the sensor was statistically significant, F(8,
10149) = 1403.79, p < 0.001 at α = 0.05. Tukey’s Honestly Significant Difference
(HSD) test for pairwise sensor effects was performed to determine statistical
significance of difference between each possible pair in the group of datasets.
For pairwise comparisons described in this section, if the p value was greater
than 0.05, the comparison was not statistically significant. Table 4.4 gives the p
values of each pairwise comparison. Tukey’s Honestly Significant Difference
(HSD) test of S2 and S7; S3 and S4; and S6 and S9 were found to be not
statistically significant at α = 0.05. The pairwise comparison of S2 and S7 was
interesting because S2 was collected from a capacitive sensor of swipe
interaction type, while S7 was collected from an optical sensor of touch
interaction type. S6 was collected using an optical touch sensor and S9 was
collected using a capacitive touch sensor. The sensor acquisition technology or
interaction type did not have an impact on similarity of minutiae count. S3 and S4
were both collected using an optical touch sensor. Results of this test were
symmetric due to the mechanics of the Tukey’s HSD test.
88
Table 4.4 p values for Tukey’s HSD test for pairwise dataset effects for Aware software minutiae count
S1
S2
S3
S4
S5
S6
S7
S8
S2
<0.05
S3
<0.05
<0.05
S4
<0.05
<0.05
>0.05
S5
<0.05
<0.05
<0.05
<0.05
S6
<0.05
<0.05
<0.05
<0.05
<0.05
S7
<0.05
>0.05
<0.05
<0.05
<0.05
<0.05
S8
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
S9
<0.05
<0.05
<0.05
<0.05
<0.05
>0.05
<0.05
<0.05
89
89
Model adequacy checking was performed by examining residual values for
minutiae count generated by MINDTCT module in NBIS. A visual analysis of the
normal probability plot, independence of samples, and constancy of variance did
not violate the assumptions for performing single factor parametric F- test. These
graphs can be found in Appendix E. Table 4.5 shows the values of M and SD for
each of the datasets. Figure 4.3 shows the boxplot of the minutiae count for all
datasets.
Table 4.5 Descriptive Statistics for MINDTCT minutiae count
Dataset
S1
S2
S3
S4
S5
S6
S7
S8
S9
N
1140
1122
1140
1128
1110
1134
1134
1128
1122
M
78.62
48.71
57.40
52.08
47.52
58.36
35.32
47.03
49.46
SD
19.58
15.61
14.48
13.36
11.02
12.64
9.08
11.54
12.01
The omnibus test for main effect of sensor dataset was statistically significant,
F(8, 10149) = 851.0, p < 0.001 at α = 0.05. Tukey’s HSD test was performed to
determine the statistical significance of difference between each possible pair in
the group of datasets. For pairwise comparisons described in this section, if the p
value was greater than 0.05, the comparison was not statistically significant.
Table 4.6 gives the p values of each pairwise comparison. Tukey’s HSD test of
S2 and S5; S2 and S8; S2 and S9; S3 and S6; and S5 and S8 were found to be
not statistically significant at α = 0.05. The interesting results were of the pairwise
comparison of S2 and S5 and S2 and S9 because S2 was collected from a
capacitive swipe sensor which S5 and S9 were collected using capacitive touch
sensors. S8 which was collected using a capacitive swipe sensor showed a
statistical similar count to every capacitive sensor except for S9. Results of this
test were symmetric.
90
Figure 4.3 Boxplot of Minutiae Count Extracted by MINDTCT
Figure 4.4 Histogram of Minutiae Count Extracted by MINDTCT
91
Table 4.6 p values for Tukey’s HSD test for pairwise dataset effects for MINDTCT minutiae count
S1
S2
S3
S4
S5
S6
S7
S8
S2
<0.05
S3
<0.05
<0.05
S4
<0.05
<0.05
<0.05
S5
<0.05
>0.05
<0.05
<0.05
S6
<0.05
<0.05
>0.05
<0.05
<0.05
S7
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
S8
<0.05
>0.05
<0.05
<0.05
>0.05
<0.05
<0.05
S9
<0.05
>0.05
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
92
92
Model adequacy checking was performed by examining residual values for
minutiae count generated by VeriFinger 5.0 extractor. A visual analysis of the
normal probability plot, independence of samples, and constancy of variance did
not violate the assumptions for performing single factor parametric F- test. Table
4.7 shows the values of M and SD for each of the datasets. Figure 4.5 shows the
boxplot of minutiae count extracted by VeriFinger 5.0.
Table 4.7 Descriptive statistics for VeriFinger 5.0 minutiae count
Dataset
S1
S2
S3
S4
S5
S6
S7
S8
S9
n
1140
1122
1140
1128
1110
1134
1134
1128
1122
M
41.72
28.32
40.25
30.74
24.38
38.62
27.53
26.11
26.15
SD
10.57
10.73
10.12
8.06
6.87
9.18
7.69
6.69
6.73
The omnibus test for main effect of sensor dataset was statistically significant,
F(8, 10149) = 685.86, p < 0.001 at α = 0.05. Tukey’s HSD test was performed to
determine the statistical significance of difference between each possible pair in
the group of datasets. For pairwise comparisons described in this section, if the p
value was greater than 0.05, the comparison was not statistically significant.
Table 4.8 gives the p values of each pairwise comparison. Tukey’s HSD test of
S8 and S8 was found to be not statistically significant at α = 0.05. S8 was
collected using a capacitive swipe sensor and S9 was collected using a
capacitive touch sensor. Results of this test were symmetric.
93
Figure 4.5 Boxplot of Minutiae Count Extracted by VeriFinger 5.0
Figure 4.6 Histogram of Minutiae Count Extracted by VeriFinger 5.0
94
Table 4.8 p values for Tukey’s HSD pairwise test for dataset effects
S1
S2
S3
S4
S5
S6
S7
S8
S2
<0.05
S3
<0.05
<0.05
S4
<0.05
<0.05
<0.05
S5
<0.05
<0.05
<0.05
<0.05
S6
<0.05
<0.05
<0.05
<0.05
<0.05
S7
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
S8
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
S9
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
>.05
95
95
4.2.2. Image Quality Analysis
Two different software were used to extract the image quality scores: Aware
Quality Software, and NFIQ. The results of these two different software were not
compared to one another.
The descriptive statistics for image quality scores generated by Aware software
are given in Table 4.9. Figure 4.7 shows the boxplot of quality scores generated
by Aware software.
Table 4.9 Descriptive statistics for Aware software quality scores
Dataset
S1
S2
S3
S4
S5
S6
S7
S8
S9
N
1140
1122
1140
1128
1110
1134
1134
1128
1122
M
44.49
79.52
68.76
72.64
79.92
71.02
75.88
76.58
68.43
Median
45
81
75
75
82
73
80
80
70
The image quality scores generated by Aware software did not satisfy the
assumptions for performing the parametric F-test. The model adequacy check of
normality of residuals did not hold. Figure 4.9 shows the normality plot of the
residuals.
96
Figure 4.7 Boxplot of Image Quality Scores Computed by Aware
Figure 4.8 Histogram of Image Quality Scores Computed by Aware
97
Figure 4.9 Normality plot of residuals
Instead an analysis of variance on rank of the response variable was performed
using the Kruskal Wallis test (Tolley, 2006). The test for main effect of the sensor
dataset was statistically significant, H (8) = 3379.77, p < 0.001 at α = 0.05.
Follow up tests were performed on pairwise comparisons of datasets to
determine which pairs of sensors were statistically significant in their differences.
Tukey’ HSD pairwise comparisons were performed on the ranks of observations
for each dataset (Tolley, 2006). Table 4.11 gives the p values of each pairwise
comparison. The test of pairwise comparison for S2 and S5; S3 and S4; S4 and
S6; S6 and S9, and S7 and S8 were found to be not statistically significant at α =
0.05. Table 4.10 shows the different acquisition technology and interaction types
of the sensors used to collect the datasets in the pairwise comparisons. The
notation in the acquisition technology and interaction type columns correspond to
the respective elements in the pairwise dataset column
98
Table 4.10 Summary of Sensor Acquisition Technologies and Interaction Types
Pairwise Dataset
(S2/S5)
(S3/S4)
(S4/S6)
(S6/S9)
(S7/S8)
Acquisition
Technology
(Capacitive/Capacitive)
(Optical/Optical)
(Optical/Optical)
(Optical/Capacitive)
(Optical/Capacitive)
Interaction Type
(Swipe/ Touch)
(Touch/Touch)
(Touch/Touch)
(Touch/Touch)
(Touch/Swipe)
Each optical touch sensor dataset was found to have similar quality scores with
dataset collected from another optical touch sensor or capacitive touch and
capacitive swipe sensor, but none of the datasets showed similarity of scores
with more than one other dataset. As a group, the optical touch sensors showed
a higher level of similarity among their datasets compared to capacitive touch
sensors. It should also be noted that the pairwise comparison results of minutiae
count from Table 4.4 for S3 and S4, and S6 and S9 were found to be similar as
well. Results of this test are symmetric.
99
Table 4.11 p values for test for difference in quality scores in every possible pairwise comparison using Aware Quality
scores
S1
S2
S3
S4
S5
S6
S7
S8
S2
<0.05
S3
<0.05
<0.05
S4
<0.05
<0.05
>0.05
S5
<0.05
>0.05
<0.05
<0.05
S6
<0.05
<0.05
<0.05
>0.05
<0.05
S7
<.001
<0.05
<0.05
<0.05
<0.05
<0.05
S8
<0.05
<0.05
<0.05
<0.05
<0.05
<0.05
>0.05
S9
<0.05
<0.05
<0.05
<0.05
<0.05
>0.05
<0.05
<0.05
100
100
Table 4.12 shows the descriptive statistics for image quality scores generated by
NFIQ. A score of one indicates highest possible quality score and five indicates
lowest possible quality score.
Table 4.12 Descriptive Statistics for NFIQ Image Quality Analysis
Dataset
S1
S2
S3
S4
S5
S6
S7
S8
S9
N
1140
1122
1140
1128
1110
1134
1134
1128
1122
M
1.29
1.91
1.75
2.00
2.18
1.77
2.03
2.14
1.58
Median
1
2
1
2
2
2
2
2
2
Figure 4.10 Boxplot of Image Quality Scores Computed by NFIQ
101
Figure 4.11 Histogram of Image Quality Scores Computed by NFIQ
NFIQ uses a 3-layer feed forward nonlinear perceptron model to predict the
image quality values based on the input feature vector of the fingerprint image
(Tabassi, Wilson, & Watson, 2004). Neural networks are non-parametric
processors, which implies that the results produced have non-parametric
characteristics (Shu-Long, Zhong-Kang, & Yan-Yan, 1991). This property of
NFIQ values precludes the use of parametric based approach for detecting
differences in quality scores between all fingerprint datasets. Instead an analysis
of variance on rank of the response variable was performed using the Kruskal
Wallis test. The test for main effect of the sensor was statistically significant,
H (8) = 1646.07, p < 0.001 at α = 0.05.
Follow up tests were performed on pairwise comparisons of sensor datasets to
determine which pairs of sensors were statistically significant in the differences of
NFIQ scores. Tukey’s HSD pairwise comparisons were performed on the ranks
of the observations for each dataset. If the p value was greater than 0.05 for any
102
of the pairwise comparisons, the comparison was not statistically significant.
Table 4.13 gives the p value of each pairwise comparison. The test of pairwise
comparison for S4 and S7; S4 and S8; S5 and S9; and S7 and S9 was found to
be not statistically significant at α = 0.05. S4 and S7 datasets were both collected
from optical sensors and touch interaction type sensors. S7 dataset was
collected from an optical touch sensor and S8 was collected from a capacitive
swipe sensor. It should be noted that S5 and S8 also showed no statistically
significant difference in minutiae count extracted by MINDTCT (Table 4.6). S8
showed a similarity of quality scores with the most number of other datasets, and
the other datasets were of both optical touch and capacitive touch types. As a
group, neither optical touch sensors nor capacitive touch sensors showed a high
level of similarity of quality scores. Results of this test are symmetric.
103
Table 4.13 p values for test for difference in quality scores in every possible pairwise comparison using NFIQ scores
S1
S2
S3
S4
S5
S6
S7
S8
S2
S3
S4
S5
<.001 <.001 <.001 <.001
<.001 <.001 <.001
<.001 <.001
<.001
S6
<.001
<.001
<.001
<.001
<.001
S7
<.001
<.001
<.001
>0.05
<.001
<.001
S8
<.001
<.001
<.001
>0.05
>0.05
<.001
>0.05
S9
<.001
<.001
<.001
<.001
<.001
<.001
<.001
<.001
104
104
4.3. Match Score Analysis
This section contains the match score results obtained using VeriFinger 5.0 and
BOZORTH3, and the subsequent analysis performed on the match score results.
4.3.1. VeriFinger 5.0 Interoperability Error Rates
The interoperability FNMR matrices are shown in Table 4.14, and Table 4.15.
The cells along the diagonal indicate enrollment and test fingerprint images from
the same fingerprint sensor. The cells off the diagonal indicate fingerprint images
from different sensors. The error rates matrix generation methodology was
described in Sections 3.10 and 3.12.2. The sensor dataset in the rows indicate
the source of the enroll sensor and the sensor dataset in the columns indicate
the source of the test sensor. All the native FNMR were found to be lower than
interoperable datasets except for the interoperable dataset {S5,S9} where S5
and S9 were both collected from a capacitive touch type sensor. The
interoperable FNMR for {S5, S9} = 0.55% and native FNMR for S5 = 1.02% as
shown in Table 4.14. The dataset for S9 had a mean minutiae count of 26.15
while dataset for S5 had a mean minutiae count of 24.38 using VeriFinger 5.0
extractor. This result indicated the interoperable dataset {S5, S9} had a larger
number of minutiae points to match compared to native dataset of S5. The
capture area of the sensor used to for S9 dataset was larger compared to
capture area used for S5 dataset. Also interesting was the extremely high FNMR
for interoperable S8 datasets. S8 dataset was collected from a capacitive sensor
of swipe interaction type and had an extremely high FNMR with datasets
collected from all different types of sensors. A cross reference analysis of the
FNMR matrix with similarity of minutiae count and quality scores from the
previous section did not show any specific relations.
105
Table 4.14 FNMR at FMR 0.01% (in percentage values)
TEST
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S8
S9
S1
0.64
8.70
7.63
6.15
7.72
44.42
15.07
100
5.19
S2
S3
S4
S5
S6
S7
S8
S9
10.61
11.61
7.29
9.32
46.14
18.99
100
4.90
5.58
6.18
5.40
9.67
17.92
11.63
99.69
3.39
8.43
0.30
1.31
1.16
5.33
2.50
100
0.24
9.30
1.46
0.23
1.03
1.54
2.42
100
1.38
13.81
1.95
1.70
1.02
6.89
3.68
100
0.73
20.62
4.56
2.13
6.54
0.17
3.54
100
2.57
13.42
2.44
1.95
3.02
2.18
1.29
100
1.19
100
100
100
100
100
100
1.30
100
7.44
0.90
1.26
0.55
2.18
2.56
100
0.11
106
106
TEST
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S8
S9
S1
0.47
6.79
5.60
4.44
5.61
31.03
11.30
99.81
3.76
S2
S3
S4
S5
S6
S7
S8
S9
7.33
8.57
5.10
5.56
33.86
14.92
100
2.87
5.05
4.78
4.44
7.64
13.42
9.89
99.63
2.96
6.49
0.24
0.95
0.85
2.99
1.73
100
0.12
7.44
1.07
0
0.85
1.30
2.12
100
0.90
10.92
1.28
1.39
0.78
4.47
2.35
100
0.49
16.18
2.73
1.60
4.42
0.17
2.71
100
1.73
10.83
1.79
1.83
2.23
1.41
0.94
100
0.77
99.81
100
100
100
100
100
1.00
100
6.11
0.54
1.08
0.42
2.15
1.85
99.93
0.11
107
107
4.3.1.1. Core Overlap
Table 4.16 summarizes the percentage of pairs used in genuine comparisons
that had cores present in both images of the pair. A higher number indicated a
larger number of pairs of images in which a core was detected.
A scatter plot (Figure 4.12) was created where the x-axis represented the
percentage of fingerprint image pairs with core overlap from each interoperable
dataset and the y-axis represented its corresponding interoperable FNMR at
fixed FMR of 0.1%. The scatter plot only contained interoperable FNMR and
interoperable dataset core overlap percentage since interoperable FNMR were
the data points of interest. An inverse relation between FNMR and percentage of
pairs with core overlap was observed. There were two data points of interest
which represent interoperable datasets {S1,S6} and {S6,S1}. The relation
between percentage of images with cores and FNMR of these two interoperable
datasets did not follow the trend observed in the other interoperable datasets,
where a higher percentage of core overlap between images was related to a
lower FNMR. S1 was collected using thermal swipe sensor and S6 was collected
using optical touch sensor. These points indicate existence of other underlying
factors which affected the interoperable FNMR, which are discussed in Section
4.7.
108
Table 4.16 Percentage of fingerprint pairs with detected core
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S8
S9
S1
75.32
65.53
78.79
78.23
72.27
78.65
75.57
75.31
78.39
S2
68.93
68.80
77.15
77.29
71.59
77.27
75.29
74.00
76.42
S3
77.16
72.79
88.77
86.89
80.82
87.23
84.05
82.60
86.58
TEST
S4
S5
76.02 71.15
71.95 67.34
86.41 81.51
86.99 81.11
80.83 77.03
86.86 80.45
83.33 78.68
81.38 76.12
84.86 79.88
S6
77.75
73.74
88.84
89.18
81.93
90.03
85.34
82.88
86.81
S7
74.08
71.62
85.06
84.69
79.37
84.63
82.71
79.22
83.63
S8
70.69
70.69
84.07
82.91
76.91
82.32
80.92
81.96
82.99
S9
77.43
72.35
87.68
86.12
79.79
86.29
83.63
82.09
86.42
109
109
Figure 4.12 Scatter plot of Consistency of Placement vs. FNMR (FMR=0.1%)
4.3.2. NBIS Interoperability Error Rates
The interoperable FNMR for match scores generated by the NBIS matcher is
shown in Table 4.17. As described in Section 3.4.2.2, a single performance
interoperability matrix was generated using a decision threshold of 40.
110
Table 4.17 FNMR (at threshold of score of 40)
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S8
S9
S1
1.71
27.22
12.56
11.25
24.12
20.78
23.78
60.89
9.63
S2
21.71
19.50
18.40
16.02
25.34
26.65
24.26
61.54
14.32
S3
10.77
23.24
2.98
4.24
10.43
8.98
10.14
60.43
3.23
TEST
S4
S5
11.07 23.04
20.78 29.59
5.62
11.29
1.95
9.18
8.24
11.06
6.18
18.86
8.78
15.29
60.13 61.54
2.93
8.27
S6
20.22
31.72
10.64
7.03
18.75
0.84
13.23
60.90
5.53
S7
21.07
27.51
9.68
6.65
13.28
11.11
8.28
61.58
5.96
S8
60.54
63.33
60.15
59.99
62.16
60.37
62.01
4.18
60.00
S9
9.98
20.51
4.69
2.72
8.42
5.85
7.49
59.85
1.28
111
111
Native FNMR for S5 was higher compared to interoperable FNMR for datasets
{S5,S3}, {S5,S4}, and {S5,S9} where S5 was collected from capacitive touch type
sensor and S3 was also collected from an optical touch type sensor, S4 was
collected from an optical touch sensor, and S9 was collected from a capacitive
touch sensor. Native FNMR for S7 was also higher compared to interoperable
FNMR for datasets {S7,S9}. S7 was collected using an optical touch sensor and
S9 was collected using a capacitive touch sensor.
Fingerprints from S8 showed a very high FNMR with interoperable datasets but
its native dataset FNMR was not the highest native dataset FNMR in the matrix.
A similar relation was observed in FNMR matrix generated using VeriFinger 5.0
matcher (Table 4.14 and Table 4.15).
4.3.3. Test of Proportions for VeriFinger 5.0 Match Scores
The test of homogeneity of proportions using the χ2 distribution was performed for
comparing FNMR of the native dataset to FNMR of interoperable dataset. FNMR
calculated at fixed FMR 0.1% was used for this test. Significance level of 0.05
was used. This test was performed for the hypothesis stated in Eq. 3.5, which
has been restated in Eq.4.1. For pairwise comparisons described in this section,
if the p value was greater than 0.05, the comparison was not statistically
significant.
H10: pi = p2 = ……..= pn
Eq. 4.1
H1A: pi ≠ p2 ≠ ……..≠ pn
For S1 dataset the FNMR was compared to all other interoperable datasets of
S1, and the test was significant at χ2 (8, N =15182) = 8938.87, p < .001. Since
the null hypothesis was rejected the Marascuillo procedure, described in Section
3.4.2.2, was used to simultaneously test differences of all interoperable datasets
with S1 native dataset. The results of the Marascuillo procedure can be
112
interpreted in the same way as Tukey’s HSD pairwise test, where the results of
the test indicate which interoperable datasets within the group are statistically
significant in their differences compared to the native dataset. The results for
pairwise comparison with S1 are shown in Table 4.19.
This overall test of homogeneity of proportions was performed one at a time for
each native dataset S2, S3, S4, S5, S6, S7, S8 and S9 and their respective
interoperable datasets. The results of the overall test are given in Table 4.18.
Table 4.18 Results of Overall Test of Proportions
Native Dataset
S2
S3
S4
S5
S6
S7
S8
S9
χ2 Statistic
χ2 (8, N =14595 ) = 8128.5
2
χ (8, N =14814 ) = 12101.87
χ2 (8, N =15075 ) = 12742
χ2 (8, N =14783 ) = 11880
χ2 (8, N =14985 ) = 9837
χ2 (8, N =15135 ) = 10919
χ2 (8, N =14987 ) = 14747
χ2 (8, N =14944 ) = 13486
p value
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
Table 4.18 showed that all tests were statistically significant in their differences
which indicated that the FNMR calculated for the interoperable matrix in Table
4.15 were all different from their respective native dataset FNMR. Since the null
hypothesis was rejected for all overall tests of proportions, the Marascuillo
procedure was used to simultaneously test the differences of all pairs. The
results for pairwise comparison for each native dataset are shown in Table 4.19.
The first cell in the rows in Table 4.19 indicated the native dataset and the
remaining cells in that row indicated its corresponding interoperable datasets. A p
value of less than 0.05 indicated a statistically significant difference. All pairwise
comparisons for the native S1 and S8 datasets were statistically different. S1
was collected using a thermal swipe sensor and S8 was collected using a
capacitive swipe sensor. S1 dataset showed a statistically significant difference
113
in all pairwise comparisons of minutiae count and image quality scores as well.
The other pairwise comparisons test showed an even distribution of similarity of
FNMR with interoperable datasets without any apparent trend among acquisition
technologies and interaction types of the sensors that the datasets were
collected from. The test of {S2, S1} was interesting since it did not show a
difference in the pairwise test of proportions but showed a difference in minutiae
count and image quality scores for all software used. S2 was collected a
capacitive swipe sensor and S7 was collected using an optical touch sensor. The
test of {S3,S4} was interesting since it did not show a difference in pairwise test
of proportions, and also did not show a difference in minutiae count and image
quality scores. S3 and S4 were both collected using optical touch sensors. These
two tests indicated that impact of minutiae count similarity and image quality
score similarity was not consistent on the pairwise test of proportions.
Interoperable datasets which contained S9 as its second dataset showed a high
level of similarity to the native datasets which were used to create the
interoperable dataset. This indicated that S9 did not degrade the performance of
an interoperable dataset compared to the performance of native datasets.
114
Table 4.19 p values of pairwise test of proportions
Native
Dataset
Pairwise Dataset for Comparison
S1
S1
S2
S3
S4
Fujtisu
S6
S7
S8
S9
>.05
<.001
<.001
<.001
<.001
<.001
<.001
<.001
S2
<.001
<.001
<.001
<.001
<.001
<.001
<.001
<.001
S3
S4
<.001 <.001
>.05 >.05
>.05
>.05
>.05 >.05
<.001 >.05
>.05 >.05
<.001 <.001
>.05 >.05
S5
<.001
<.001
>.05
>.05
<.001
>.05
<.001
>.05
S6
<.001
<.001
<.001
<.001
<.001
>.05
<.001
<.001
S7
<.001
<.001
<.001
<.001
>.05
>.05
<.001
>.05
S8
<.001
<.001
<.001
<.001
<.001
<.001
<.001
S9
<.001
>.05
>.05
>.05
>.05
>.05
>.05
<.001
<.001
115
115
4.3.4. Test of Proportions for BOZORTH3 Match Scores
The test of homogeneity of proportions using the χ2 distribution was performed for
FNMR generated using BOZORTH3. Significance level of 0.05 was used. This
test was performed for the hypothesis stated in Eq. 3.5, which was restated in
Eq. 4.2. For pairwise comparisons described in this section, if the p value was
greater than 0.05, the comparison was not statistically significant.
H10: pi = p2 = ……..= pn
Eq. 4.2
H1A: pi ≠ p2 ≠ ……..≠ pn
For S1 dataset, the test was significant at the χ2 (8, N =15192) = 4874,
p < .001. Since the null hypothesis was rejected, the Marascuillo procedure,
described in Section 3.4.2.2, was used to simultaneously test differences of all
interoperable datasets with S1. The results of the Marascuillo procedure can be
interpreted in the same way as Tukey’s HSD pairwise test, where the results of
the test indicate which specific pairs within the group are statistically significant in
their differences. The results for pairwise comparison with S1 are shown in Table
4.21.
The overall test of homogeneity of proportions was performed one at a time for
each native dataset S2, S3, S4, S5, S6, S7, S8 and S9 and their respective
interoperable datasets. The results of the overall test are given in Table 4.20.
116
Table 4.20 Results of overall test of proportions
Native Dataset
S2
S3
S4
S5
S6
S7
S8
S9
χ2 Statistic
χ (8, N =15021 ) =2921
χ2 (8, N =15237 ) = 6517
χ2 (8, N =15086 ) = 7742
χ2 (8, N =14886) = 5270
χ2 (8, N =15165 ) = 5941
χ2 (8, N =15174 ) = 5595
χ2 (8, N =15102 ) = 12755
χ2 (8, N =15008 ) = 8326
2
p value
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
Table 4.21 showed that all tests were statistically significant in their differences
which indicated that the FNMR calculated for the interoperability matrix in Table
4.17 were all different from their respective native dataset FNMR. Since the null
hypothesis was rejected for all overall tests of proportions, the Marascuillo
procedure was used to simultaneously test the differences of all pairs. The
results for pairwise comparison for each native dataset are shown in Table 4.21.
The first cell in the rows in Table 4.21 indicated the native dataset and the
remaining cells in that row indicated its corresponding interoperable datasets. S5
and S7 exhibited the highest number of interoperable datasets with FNMR similar
to their respective native datasets. S5 was collected using a capacitive touch
sensor and S7 was collected using an optical touch sensor. The pairwise
comparison of interoperable {S5,S3}, {S5,S4}, {S5, S9} with S5, and pairwise
comparisons of interoperable {S7,S3}, {S7,S4}, {S7,S9} with S9 were not
statistically different. The interoperable datasets created with S3, S4, and S9
were common to both tests. S3 and S4 were both optical touch sensors and S9
was a capacitive touch sensor.
117
Table 4.21 p values of pairwise test of proportions
Native Sensor
S1
S1
S2
S3
S4
S5
S6
S7
S8
S9
<.001
<.001
<.001
<.001
<.001
<.001
<.001
<.001
S2
<.001
<.001
<.001
<.001
<.001
<.001
<.001
<.001
Pairwise Sensor for Comparison
S3
S4
S5
S6
S7
<.001 <.001 <.001 <.001 <.001
<.001 <.001 <.001 <.001 <.001
<.001 <.001 <.001 <.001
<.001
<.001 <.001 <.001
>.05 >.05
<.001 <.001
<.001 <.001 <.001
<.001
>.05 >.05 <.001 <.001
<.001 <.001 <.001 <.001 <.001
<.001 <.001 <.001 <.001 <.001
S8
S9
<.001 <.001
<.001 <.001
<.001 >.05
<.001 >.05
<.001 >.05
<.001 <.001
<.001 >.05
<.001
<.001
118
118
4.3.5. Test of Similarity of VeriFinger 5.0 Match Scores
This section contains results for test of similarity of the genuine and imposter
match scores generated by VeriFinger 5.0 for native and interoperable datasets.
These match scores were the raw scores computed by VeriFinger 5.0. The test
was performed by running nine different similarity tests. For each test, one native
dataset was chosen as the control dataset and the rest of its interoperable
datasets were compared to the control dataset.
The first test of similarity was performed by choosing S1 dataset as the native
dataset and comparing its raw genuine match scores to the corresponding raw
match scores from the other eight interoperable datasets. The residuals of
genuine match scores generated for this test did not satisfy the model adequacy
tests for performing a parametric F-test. Figure 4.13 shows the normality plot of
residuals for S1 dataset genuine match scores and its corresponding
interoperable dataset genuine match scores.
Figure 4.13 Normality plot of residuals of S1 genuine match scores
The Kruskal-Wallis test for main effect of the sensor was found to be statistically
significant, H (8) = 4009, p < 0.001 at α = 0.05. Follow up tests performed on
119
pairwise comparison of every interoperable dataset with the native dataset
showed that the test was significant with p < 0.001 at α = 0.05. The same tests
were performed one at a time for remaining native datasets. The normality plots
for other datasets can be found in Appendix H. Table 4.22 shows the results of
all tests. The p value for each test showed that the differences were statistically
significant and the null hypothesis was rejected. Follow up tests were performed
by comparing each pair of interoperable dataset with its corresponding native
dataset. All pairwise comparisons were found to be statistically significant in their
differences.
Table 4.22 Results of overall test of similarity of genuine match scores
Native Dataset
S2
S3
S4
S5
S6
S7
S8
S9
Kruskal Wallis Test
H (8) = 2209
H (8) = 9716
H (8) = 9476
H (8) = 4128
H (8) = 4684
H (8) = 4111
H (8) = 2369
H (8) = 4885
p value
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
A test of similarity was performed on raw match scores of imposter comparisons
for S1 dataset and its corresponding interoperable datasets. The raw match
scores generated for this test did not satisfy the model adequacy tests for
performing a parametric F-test. The normality plot is shown in Figure 4.14. The
Kruskal-Wallis test for main effect of the sensor was found to be statistically
significant, H (8) = 80685, p < 0.001 at α = 0.05. Follow up tests were performed
on pairwise comparison of every S1 interoperable dataset with the S1 native
dataset showed that the test was statistically significant with p < 0.001 at
α = 0.05. The tests did not show any similarity of raw match scores of genuine
comparisons between native and interoperable datasets.
120
Figure 4.14 Normality plot of residuals of S1 imposter match scores
The same tests were performed one at a time for remaining native datasets. The
normality plots for other datasets can be found in Appendix G. Table 4.23 shows
the results of the tests. The p value for each test showed that the differences
were statistically significant and the null hypothesis was rejected for each test.
The pairwise comparison for each native dataset was also found to be
statistically significant. The tests did not show a similarity of raw match scores of
imposter comparisons between native and interoperable datasets.
Table 4.23 Results of overall test of similarity of imposter match scores
Native Dataset
S2
S3
S4
S5
S6
S7
S8
S9
Kruskal Wallis Test
H (8) = 71246
H (8) = 99085
H (8) = 112211
H (8) = 122373
H (8) = 163290
H (8) = 161126
H (8) = 240592
H (8) =129842
p value
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
121
4.4. Impact of Quality on Interoperability
This section describes the impact of removing low quality fingerprint images and
recalculating the interoperable FNMR. The fingerprint images which had NFIQ
quality scores of four and five were not used for the matching operation. NFIQ
provides five categories of quality scores and experiments have been performed
by NIST to predict the impact of NFIQ quality scores on matching operations.
Image quality scores from Aware Quality Software provided scores on a scale of
1-100 and would have required an arbitrary decision threshold for removal of
fingerprint images. The NIST MINEX test report indicated that a reduction in
interoperability FNMR would be observed since poorly performing images were
not used (Grother et al., 2006). The analysis performed in this section examined
the impact of removing low quality images on variance of interoperability FNMR.
VeriFinger 5.0 template generator and matcher were used to perform the
matching operations. Note that dataset S8 was not included in this test since
there was no impact of changing the fixed FMR operational points on the
interoperable FNMR and the average quality of the fingerprint images from S8
was not the worst of all the native datasets. Section 4.7 analyzes the potential
factors which affected the interoperable FNMR for S8. Table 4.24 shows the
number of genuine comparisons performed on full datasets and Table 4.25
shows the number of genuine comparisons performed on datasets with
fingerprint images of NFIQ scores of less than four.
122
Table 4.24 Number of comparisons in full datasets
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S9
S1
1698
1677
1679
1686
1652
1701
1695
1671
S2
1677
1683
1650
1665
1623
1668
1668
1650
S3
1677
1647
1666
1671
1632
1536
1674
1647
TEST
S4
1689
1665
1671
1692
1644
1683
1692
1665
S5
1657
1629
1634
1647
1655
1653
1656
1631
S6
1695
1668
1533
1683
1650
1701
1692
1668
S7
1698
1311
1674
1692
1653
1692
1701
1674
S9
1675
1653
1649
1665
1628
1671
1674
1680
123
123
Table 4.25 Number of comparisons for datasets with images of NFIQ < 4
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S9
S1
1655
1639
1628
1645
1626
1651
1654
1667
S2
1636
1660
1595
1616
1590
1618
1621
1644
S3
1652
1626
1624
1635
1620
1504
1643
1647
TEST
S4
1659
1635
1620
1650
1617
1647
1651
1662
S5
1623
1599
1593
1612
1627
1620
1626
1626
S6
1668
1644
1500
1653
1635
1648
1666
1665
S7
1668
1293
1633
1653
1635
1662
1667
1671
S9
1647
1635
1611
1632
1614
1641
1644
1660
124
124
E
N
R
O
L
L
S1
S2
S3
S4
S5
S6
S7
S9
S1
0.001
0.055
0.065
0.037
0.046
0.306
0.119
0.025
S2
0.050
0.046
0.033
0.036
0.066
0.116
0.084
0.026
S3
0.038
0.057
0
0.007
0.005
0.019
0.007
0
TEST
S4
0.033
0.060
0.008
0.001
0.006
0.011
0.007
0.007
S5
0.038
0.098
0.006
0.01
0.004
0.032
0.012
0.004
S6
0.267
0.150
0.022
0.014
0.040
0
0.019
0.013
S7
0.097
0.098
0.007
0.011
0.015
0.012
0.004
0.007
S9
0.028
0.051
0.002
0.007
0.003
0.012
0.007
0.001
125
125
Table 4.26 shows the interoperability FNMR matrix for the reduced datasets at
fixed FMR of 0.1%. Note that the interoperable FNMR for {S1,S6} increased to
0.267% compared to the full dataset interoperability FNMR of 0.25%. S1 dataset
was collected using a thermal swipe sensor and S6 dataset was collected using
an optical touch sensor. This is interesting since all the other FNMR reduced
compared to FNMR calculated for full datasets.
The tests performed in Section 4.4.3 were repeated on interoperable fingerprint
datasets which contained fingerprint images with NFIQ score of one, two or
three. Table 4.27 shows the results for overall test of proportions performed on
each native dataset. All the tests were found to be significant in their differences
at α = 0.05.
Table 4.27 Results of overall test of proportions
Native Dataset
S1
S2
S3
S4
S5
S6
S7
S9
χ2 Statistic
χ2 (8, N =14793 ) = 9231.49
χ2 (8, N =14039 ) = 8233.23
χ2 (8, N =14356 ) = 12297.23
χ2 (8, N =14666 ) = 12755.65
χ2 (8, N =14543 ) = 11923.89
χ2 (8, N =14569 ) = 9819.06
χ2 (8, N =14753 ) = 11380.38
χ2 (8, N =14847 ) = 13461.69
p value
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
p < 0.001
The results from pairwise comparisons are shown in Table 4.28. The
recalculated pairwise comparisons showed that the test of proportions for
{S3,S7} dataset did not show any statistically significant difference while in
Section 4.4.3 the test of proportion for the datasets showed a statistically
significant difference. S3 and S7 both collected using optical touch sensors. This
test indicated that removing the low quality images reduced the difference of
variance of FNMR between these datasets.
126
The recalculated results for {S6,S4}, {S6,S7}, and {S6,S9} datasets showed a
statistically significant difference. The results from Section 4.4.3 of the same test
showed no statistically significant difference between the native and
interoperable datasets. This result was interesting as it indicated that difference
of variance of FNMR increased for these interoperable datasets when the lowest
NFIQ score images were removed. It should also be noted that S6, S4, and S7
were all optical touch sensors. Removal of low quality images did not have a
consistent effect on FNMR of interoperable datasets which were all collected
using optical touch sensors.
The rows in Table 4.28 indicate the native sensor dataset and each row indicates
the results of the pairwise comparison with remaining sensor datasets with the
native dataset.
127
Table 4.28 Results of pairwise test of proportions
Native
Sensor
Pairwise Sensor for Comparison
S1
S1
S2
S3
S4
Fujtisu
S6
S7
S9
>.05
<.001
<.001
<.001
<.001
<.001
<.001
S2
<.001
<.001
<.001
<.001
<.001
<.001
<.001
S3
<.001
>.05
>.05
>.05
<.001
>.05
>.05
S4
<.001
>.05
>.05
>.05
<.001
>.05
>.05
S5
<.001
<.001
>.05
>.05
<.001
>.05
>.05
S6
<.001
<.001
<.001
<.001
<.001
>.05
<.001
S7
<.001
<.001
>.05
>.05
>.05
<.001
S9
<.001
>.05
>.05
>.05
>.05
<.001
>.05
>.05
128
128
4.5. Acquisition and Interaction Level Interoperability
A follow up evaluation of interoperability was performed by grouping the datasets
into three categories: datasets collected using swipe interaction type sensors,
datasets collected using optical touch type sensors, and datasets collected using
capacitive touch type sensors. The evaluation of different groups allowed for
examination of interoperability at the acquisition and interaction level, not the
sensor level. S1 and S2 were placed in the Swipe group. S3, S4, S6 and S7
were placed in the Optical Touch group. S5 and S9 were placed in the Capacitive
Touch group. S8 dataset was excluded from this analysis as it showed an
extremely high interoperability FNMR with the rest of the datasets (Table 4.15).
VeriFinger 5.0 was used to generate the interoperability FNMR matrix at fixed
FMR of 0.1%.
Swipe
4.81
E
Swipe
N
R
Optical Touch
11.93
O
L Capacitive Touch 4.76
L
TEST
Optical Touch Capacitive Touch
11.75
6.58
1.53
1.89
1.47
0.45
The {Capacitive Touch, Optical Touch} interoperable dataset showed the lowest
interoperable FNMR indicating a high level of interoperability between optical
touch and capacitive touch technologies. The {Capacitive Touch, Swipe}
interoperable dataset had a lower FNMR than the Swipe native dataset which
indicated that the interoperable dataset performed better than the native dataset.
The {Swipe, Optical Touch} dataset had the highest FNMR and indicated the
lowest degree of interoperability. The results showed that Capacitive Touch
dataset had the lowest interoperable FNMR compared to Optical Touch and
129
Swipe interoperable datasets. The interoperable datasets generated with S9
dataset showed a high level of similarity of FNMR with the other native datasets.
Since S9 was part of the Capacitive Touch group it showed a better interoperable
FNMR with other groups. In combination with the sensor level interoperability
analysis, these results show the effect of interoperability at the acquisition and
interaction level.
4.6. Investigative Analysis
This section analyzed the impact of difference in quality scores, ridge bifurcation
count and ridge ending count on the match scores, and the results were grouped
by acquisition technology. The matching results of interoperable datasets with S8
were not used for this analysis since previous analysis did not indicate quality or
minutiae count being responsible for the extremely high FNMR. A correlation
matrix was generated for examining the relation between match score of a pair of
fingerprint images, the difference in quality scores for the pair of fingerprint
images, the difference in ridge bifurcation count of the pair of images, and the
difference in ridge ending count of the pair of images (Table 4.30).
Table 4.30 Correlation matrix of match score, quality score difference, ridge
bifurcation count difference, ridge ending count difference
Quality Score
Match
Quality Score
Ridge Bifurcation
Score
Difference
Count Difference
0.16
Difference
Ridge Bifurcation
-0.15
0.40
0.11
0.52
Count Difference
Ridge Ending
Count Difference
0.32
130
A scatter plot of difference in Aware quality scores between the pair of images
and the match scores was generated. The match scores were the raw match
scores generated by VeriFinger 5.0 for genuine comparisons of native and
interoperable datasets. The scatter plot is grouped by the acquisition technology
used to capture the two fingerprint images that were matched. A higher
difference in quality scores was related to a lower match score. The Same Type
and Optical-Capacitive groupings showed a relatively high match score with high
difference in match scores. The grouping of data points showed that ThermalCapacitive had relatively lower match scores compared to Optical –Capacitive
and Same Type groupings. The Thermal-Capacitive and Optical-Thermal
groupings did not show any specific relation between difference in quality scores
and match scores, which indicated that difference in quality scores and its
relation to match scores was impacted by the acquisition technologies.
Figure 4.15 Scatter plot of VeriFinger 5.0 Match Score vs. Difference in Quality
Scores
131
A scatter plot of match score vs. difference in ridge bifurcation count was
generated and grouped by acquisition technologies used to capture the pair of
fingerprints that were matched. A lower difference in ridge bifurcation count
indicated that a similar number of ridge bifurcation points were detected in the
two fingerprints being compared. The scatter plot showed that difference
between ridge bifurcation count and the match score showed a negative relation
for Same Type and Optical-Capacitive groupings. The difference in ridge
bifurcation count and match score did not show a specific relation for the OpticalThermal and Thermal-Capacitive groupings.
Bifurcation Count
A scatter plot of match score vs. difference in ridge ending count was generated
and grouped by acquisition technologies used to capture the pair of fingerprints
that were matched. A lower difference in ridge ending count indicated that a
similar number of ridge ending points were detected in the two fingerprints being
132
compared. A similar relation to difference in ridge bifurcation count was observed
in this scatter plot. The difference in ridge ending count and match score did not
show a specific relation for Thermal –Capacitive and Optical-Thermal groupings.
The Same Type and Optical –Capacitive groupings showed a stronger relation
between the two variables.
Ending
The three scatter plots showed that images from thermal sensor when compared
to optical and capacitive sensors did not follow the relation observed for the
fingerprint images from the other groups. This makes prediction of interoperability
between thermal and other types of sensors harder compared to interoperability
between capacitive and optical sensors. These results also indicated a need to
perform further analysis on interoperability match scores for fingerprint images
collected from thermal sensors.
133
4.7. Fingerprint Image Transformation
It was observed that fingerprint images of native S8 dataset had FNMR very
similar to other native datasets, but its interoperable dataset had lowest FNMR of
99.63%. Fingerprint for S8 dataset were collected using a capacitive swipe
sensor. The image quality scores for S8 dataset did not show an extreme
deviation from image quality scores of other datasets and the same was
observed for minutiae count. The consistency of placement measure in Section
4.4.1.1 also did not show a deviation from other interoperable datasets. These
factors indicated that the even though similar minutiae were being detected
between images of S8 dataset and other datasets, the placement of minutiae in
the fingerprint images were different enough for the matcher to incorrectly reject
a genuine comparison. Further evaluation of the raw captured images showed
that distance between ridge lines for S8 dataset images was shorter compared to
all other fingerprint datasets. Figure 4.18 shows a skeletonized fingerprint image
of the same finger from dataset S3 and S8. S3 was collected using an optical
touch sensor.
134
Figure 4.18 Skeletonized fingerprint image from S8(left) and S3(right)
A ridge spacing profile for each image was created using the IMPROFILE image
processing method in MATLABTM. The ridge spacing profile was created by
traversing a distance of 100 pixels starting from the core and moving along the yaxis one pixel at a time. The IMPROFILE method computes the intensity values
along a traversal line in an image. The 100 pixel distance was chosen as it would
allow the IMPROFILE method to intersect sufficient ridge lines to get an
understanding of the distance between successive ridge lines. Figure 4.19 shows
the traversal path.
135
Figure 4.19 Traversal path for IMPROFILE
Figure 4.20 shows output from the IMPROFILE procedure for a single image
from S8 dataset, where a downward spike in the graph indicated the intersection
of the traversal route with a ridge line. Figure 4.21 shows output from the
IMPROFILE procedure for a single image from S3 dataset. Both images from S3
and S8 dataset were of the same finger. Comparison of the two graphs indicated
that distance between successive ridges for S3 image was significantly larger
than distance between successive ridges for S8 image. In order to make the
distance between successive ridges more consistent between the two images, all
images for S8 dataset were transformed using 2-D affine spatial transformation
process. The transformation parameters included scaling along the x-axis and yaxis. Figure 4.22 shows the image after transformation. It should be noted that
the scaling parameters were approximated using only one pair of image, and
then applied to all the images of S8 dataset.
136
Figure 4.20 IMPROFILE output for S8 dataset image
Figure 4.21 IMPROFILE output for S3 dataset image
137
Figure 4.22 Fingerprint image after transformation
The transformed dataset was relabeled to S8’. Using images from S8’, native and
interoperable FNMR were recalculated using VeriFinger 5.0. Table 4.31 shows
the FNMR at fixed FMR of 0.1%.
138
Table 4.31 Recalculated FNMR (FMR = 0.1%)
S8’
S1
S2
S3
S4
S5
S6
S7
S8’
S9
11.81
8.84
6.43
6.70
12.16
21.68
9.94
0.65
6.57
Table 4.32 FNMR for S3 dataset (FMR=0.1%)
S3
S1
S2
S3
S4
S5
S6
S7
S9
8.57
4.78
0.24
1.07
1.28
2.73
1.79
0.54
139
139
The FNMR for interoperable S8’ datasets showed a significant reduction, and
also a decrease in native dataset FNMR. It is worth noting that interoperable
FNMR of datasets {S8’,S4} and {S8’,S9} were comparable with FNMR of
{S8’,S3}. Comparing these results with FNMR in Table 4.32, which were have
been restated from Table 4.15, it was observed that interoperable FNMR for S8’
were lower for datasets which had a lower interoperable FNMR with S3,
specifically {S3,S4} and {S3,S9}. This indicated that by transforming images of
S8 dataset to be more similar to S3 dataset, there was an associated effect of
reducing interoperable FNMR of {S8’,S4} and {S8’,S9}.
The datasets {S1,S6} and {S6,S1} also exhibited relatively high interoperable
FNMR of 31.03% and 33.86% respectively, using VeriFinger 5.0 matcher. The
ridge spacing profile was computed for a single fingerprint image from S1 and S6
dataset along the x-axis and the y-axis. Figure 4.23 shows the ridge spacing
profile along the x-axis for S1 and S6, and Figure 4.24 shows the ridge spacing
profile along the y-axis for S1 and S6.
Figure 4.23 Ridge spacing profile along x-axis. S1 image on left and S6 image on
right
140
Figure 4.24 Ridge spacing profile along y-axis. S1 mage on left and S6 image on
right
Figure 4.23 showed that S1 ridge spacing profile along the x-axis had a larger
distance between successive ridges compared to ridge spacing profile for S6. An
opposite effect was observed in the ridge spacing profile along the y-axis for the
S1 and S6 dataset images. S1 was collected using a thermal swipe sensor and
S6 was collected using an optical touch sensor. The difference in interaction
between the two sensors was a potential factor in the difference in ridge spacing
between the two images. The swipe action introduced an elastic deformation of
the finger skin which lead to an increased space between ridges and contributed
to a higher FNMR for {S1, S6} and {S6, S1} datasets.
Reliability of the image transformation method was entirely dependent on the
ability to detect a core in both the fingerprint images. This was necessary as a
common anchor point for the two fingerprint images was required as the starting
point of the IMPROFILE method. This method would be efficient only if used in a
1:1 matching, or verification mode. The application of the transformation
methodology described in this section was performed at a sensor level since only
a pair of images was used to approximate the transformation scaling parameters
for the entire dataset collected using the specific sensors. This method can be
141
modified where a ridge spacing profile is created every time a pair of images
need to be matched, and then generate the scaling parameters based on the
ridge spacing profile. Such a method would make it a sensor agnostic method of
transforming images and would have wider applicability. The results in this
section provided justification for further development and analysis of transforming
images using the ridge spacing profile.
142
CHAPTER 5. CONCLUSIONS AND RECOMMENDATIONS
This dissertation has described the design, formulation and application of a
statistical analysis framework for testing interoperability of fingerprint sensors and
its effect on system performance. This chapter reviews the work presented in this
dissertation, highlights the contributions made by this work and provides
recommendations for future work.
5.1. Conclusions
•
Minutiae count similarity of fingerprint images collected from different
sensors did not show a relation to a specific acquisition technology or
interaction type. There were no common pairwise tests which were
statistically similar for minutiae count datasets extracted using Aware,
MINDTCT, and VeriFinger 5.0.
•
Fingerprint images collected from optical touch sensors showed a higher
level of similarity in quality scores with fingerprints collected from other
optical touch sensors.
•
Performance of minutiae based matchers was significantly affected if the
pair of fingerprints being matched were captured from different sensors.
•
Similarity of minutiae count and image quality scores did not have an
impact on similarity of FNMR for native and interoperable datasets.
•
Higher minutiae count for a particular dataset did not have an impact on
FNMR of its corresponding interoperable datasets.
143
•
Performance of interoperable datasets cannot be predicted by separately
analyzing performance of native datasets.
•
Test of similarity of raw match scores between native and interoperable
datasets did not provide remarkable results for the raw match scores
provided by VeriFinger 5.0 and BOZORTH3 matchers. This is not a
deficiency of the test, but indicated a need for transforming the scores
over a range which provides a normal distribution.
•
Interoperable datasets which had a higher percentage of pairs of
fingerprint images in which a core was detected had a positive relation
with a lower FNMR, with the only exception of {S1,S6} and {S6,S1}
interoperable datasets. Consistent interaction of the finger with a sensor is
an important factor in improving performance, and it becomes even more
important when the fingerprints are captured from different sensors. A
simple metric for consistent placement is required for collecting
fingerprints from different sensors, and this dissertation showed that core
overlap provides a metric which has an impact on FNMR of interoperable
datasets.
•
Removing low quality images from interoperable datasets did not lead to a
reduction in statistical variance of FNMR for interoperable datasets,
although the absolute FNMR was reduced for all native and interoperable
datasets.
•
Consistency of ridge spacing between interoperable datasets was a very
important factor in reducing FNMR of interoperable datasets.
•
The Capacitive Touch sensor group showed the lowest interoperable
FNMR with Swipe and Optical Touch sensor groups.
•
A negative relation between match scores and ridge ending count and
ridge bifurcation count was observed for fingerprints collected from the
same type of sensors and optical and capacitive sensors. Matching of
fingerprints collected from thermal sensor and optical sensors, and
144
thermal sensor and capacitive sensors did not show a relation between
those variables.
5.2. Contributions
•
Dataset of fingerprints from 190 individuals on nine different fingerprint
sensors in a controlled environment which can be used in testing
interoperability of feature extractors and feature matchers.
•
Collection of finger skin characteristics and pressure placed on the sensor.
•
Formulation of a framework for statistically testing interoperability of
fingerprint sensors in terms of FNMR.
•
Application of analysis framework to data collected in a controlled
experiment.
•
Enhanced enrollment procedures for reducing FNMR of interoperable
datasets.
•
Ability to statistically test similarity of FNMR of native dataset with
interoperable dataset.
•
A sensor agnostic transformation process to reduce FNMR of
interoperable datasets.
5.3. Future Work
Since this dissertation limited itself to a verification scenario, performance was
solely measured in terms of FNMR. The dataset collected as part of this
dissertation can be used to analyze False Match Rates (FMR) if an identification
scenario is of interest. The statistical analysis framework can be modified to test
for FMR of interoperable datasets, and it would be interesting to understand the
impact of interoperability on FMR of fingerprint datasets collected from multiple
sensors.
145
The impact of removing low quality images from interoperable datasets did not
lead to a higher level of similarity in FNMR between interoperable datasets and
native datasets. The level of similarity was used as a measure of variance
between interoperable and native datasets FNMR. The results indicated that
further work is required to investigate the impact of quality on the variance of
interoperable datasets performance. The relation between impact of quality and
the types of sensors which were used to create the interoperable datasets needs
to be analyzed more closely. The FBI has created image quality specifications to
measure the fidelity of the fingerprint sensor. The impact of the different
components of the image quality specifications on the interoperable error rates
would be a relevant research.
The relationships between image quality, temperature, moisture content, oiliness
and elasticity were too complex for fitting linear models. Although separate
analysis of the relations between quality score and finger skin characteristics
were performed, analysis which examines these interactions together would be
interesting. Higher order models or neural networks would be more appropriate
to understand the relation between skin characteristics and image quality.
Understanding the impact of the physiological skin factors on quality is important
since improving quality can reduce FNMR of interoperable datasets.
The dataset collected for this dissertation can be used for evaluating
interoperability of sensors, feature extractors, and feature matchers as part of the
same experiment. This dissertation did not analyze interoperability of feature
extractors and feature matchers, and previous studies related to interoperability
did not analyze the effect of interoperability of sensors. A combined analysis of
fingerprint sensors, feature extractors and feature matchers would be beneficial.
The image transformation model described in Section 4.7 indicated the need to
make the spacing between ridgelines in the pair of fingerprint images more
146
consistent in order to reduce FNMR of interoperable datasets. The method
described in this dissertation used a linear transformation to make the spacing
between ridges more uniform. This is a simple model which works if distortion
between the two fingerprint images is relatively consistent, but only differing in
scale. This model can be extended to non linear transformation models by
creating a ridge spacing profile along eight quadrants as shown in Figure 5.1. A
corresponding set of points extracted along different directions of the fingerprint
images can be used to generate transformation models where the distortion
varies only in certain parts of the fingerprint image.
Figure 5.1 Eight quadrant ridge spacing profile
The distortion of fingerprint images is heavily affected by the type of sensor and
the habituation of users with the sensor. Further work into transforming the image
so that the distortion of fingerprint images would be reduced without having any
a-priori knowledge about the fingerprint sensors would be of immense interest to
the field of fingerprint recognition. Sensor agnostic transformation methods, even
if limited in its capabilities, would significantly augment the current methodologies
being investigated.
147
LIST OF REFERENCES
Abdurrachim, D., Kunieda, H., Li, D., & Qi, J. (2005). A hybrid method for
fingerprint image quality calculation. Paper presented at the Automatic
Identification Advanced Technologies, Buffalo, N.Y.
Adhami, R., & Meenen, P. (2001). Fingerprinting for security. IEEE Potentials,
20(3), 33-38.
Aguilar, J., Fernandez, F., & Garcia, J. (2005). A Review of Schemes for
Fingerprint Image Quality Computation. Paper presented at the 3rd COST
275 Workshop Biometrics on the Internet.
Allen, R., Prabhakar, S., & Sankar, P. (2005). Fingerprint Identification
Technology. In J. Wayman, A. Jain, D. Maltoni & D. Maio (Eds.), Biometric
Systems (pp. 21-61). London: Springer Verlag.
Alonso-Fernandez, F., Veldhuis, R. N. J., Bazen, A. M., Fierrez-Aguilar, J., &
Ortega-Garcia, J. (2006, May). On the relation between biometric quality
and user-dependent score distributions in fingerprint verification. Paper
presented at the Workshop on Multimodal User Authentication, Toulouse,
France.
Amengual, J. C., Perez, J. C., Prat, F., Saez, S., & Vilar, J. M. (1997, 14th July).
Real Time Minutiae Extraction in Fingerprint Images. Paper presented at
the 6th International Conference on Image Processing and its
Applications, Dublin.
Amin, A., & Yager, N. (2004). Fingerprint Verification based on Minutiae
Features: a review. Pattern Analysis Application, 7, 94-113.
Ashbourn, J. (2000). Biometrics: Advanced Identitiy Verification. London:
Springer Verlag.
Bazin, A., & Mansfield, T. (2007, 7-8 June). An Investigation of Minutiae
Template Interoperability. Paper presented at the AutoID 2007, Alghero,
Italy.
Bigun, J., Fronthaler, H., & Kollreider, K. (2006). Automatic Image Quality
Assessment with Application in Biometrics. Paper presented at the
Computer Vision and Pattern Recognition Workshop, New York.
Biometric Sample Quality Standard - Part 1: Framework. (2006).): ISO/IEC
JTC1/SC37.
148
Boashash, B., Deriche, M., & Kasaei, S. (1997). Fingerprint Feature
Enhancement using Block-Direction on Reconstructed Images. Paper
presented at the International Conference on Information,
Communications, and Signal Processing, Singapore.
Bolle, R., & Ratha, N. K. (1998, 16-20 August). Effect of Controlled Image
Acquisition on Fingerprint Matching. Paper presented at the 14th
International Conference on Pattern Recognition, Brisbane, Australia.
Bolle, R. M., Connell, J. H., Pankanti, S., Ratha, N. K., & Senior, A. W. (2004a).
Basic System Errors. In Guide to Biometrics (pp. 61-86). New York:
Springer Verlag.
Bolle, R. M., Connell, J. H., Pankanti, S., Ratha, N. K., & Senior, A. W. (2004b).
Performance Testing. In Guide to Biometrics (pp. 105-128). New York:
Springer-Verlag.
Campbell, J., & Madden, M. (2006). ILO Seafarers' Identity Documents Biometric
Interoperability Test Report Geneva: International Labour Organization.
Chan, K., Huang, Z., & Liu, J. (2000). Direct Minutiae Extraction from Gray-Level
Fingerprint Image by Relationship Examination. Paper presented at the
International Conference on Image Processing, Vancouver, Canada.
Chen, S., Jain, A. K., Karu, K., & Ratha, N. K. (1996). A Real-Time Matching
System for Large Fingerprint Database. IEEE Transactions on Pattern
Analysis and Machine Intelligence, 18(8), 799-813.
Clear Registered Traveler. (2006). Retrieved January 31, 2007, from
HTTP://WWW.FLYCLEAR.COM/INDEX.HTML
Cole, S. A. (2001). Suspect Identities: A History of Fingerprinting and Criminal
Identification. Cambridge, Massachusetts: Harvard University Press.
Cole, S. A. (2004). History of Fingerprint Pattern Recognition. In N. Ratha & R.
Bolle (Eds.), Automatic Fingerprint Recognition Systems (pp. 1-25). New
York: Springer.
Daugman, J., & Williams, G. O. (1996). A proposed standard for biometric
decidability. Paper presented at the CardTech SecureTech, Atlanta, GA.
Dean, B., Salstrom, R., & Wayman, J. (2001). Studies of Fingerprint Imaging
Systems in Social Service Applications. Paper presented at the
International Conference on Management of Engineering and Technology
Portland.
DigitalPersona. (2006). DigitalPersona Pro. Retrieved January 31, 2007, from
HTTP://WWW.DIGITALPERSONA.COM/RESOURCES/DOWNLOADS/P
RO%20V4_010-06.PDF
Doddington, G., Kamm, T., Martin, A., Ordowski, M., & Przybocki, M. (1997). The
DET curve in assessment of detection task performance. Paper presented
at the Eurospeech, Greece.
Elliott, S. J., Kukula, E., & Modi, S. (2007). Issues Involving the Human Biometric
Sensor Interface. In S. Yanushkevich, P. Wang & S. Srihari (Eds.), Image
Pattern Recognition Synthesis and Analysis in Biometrics (Vol. 67, pp.
400): World Scientific Publishers.
149
Elliott, S. J., & Modi, S. K. (2006). Impact of Image Quality on Performance:
Comparison of Young and Elderly Fingerprints. Paper presented at the 6th
International Conference on Recent Advances in Soft Computing (RASC),
Canterbury, UK.
Elliott, S. J., & Sickler, N. C. (2005). An evalulation of fingeprint image quality
across an elderly population vis-avis an 18-25 year old population. Paper
presented at the International Carnahan Conference on Security
Technology, Las Palmas, Gran Canaria.
Elliott, S. J., & Young, M. (2007). Image Quality and Performance Based on
Henry Classification and Finger Location. Paper presented at the AutoID
2007, Alghero, Italy.
Engeler, W., Frank, P., & Leung, M. (1990, 24th September). Fingerprint Image
Processing Using Neural Network. Paper presented at the IEEE
Conference on Computer and Communication Systems, Hong Kong.
Eshera, M., & Shen, W. (2004). Feature Extraction in Fingerprint Images. In R.
Bolle & N. Ratha (Eds.), Automatic Fingerprint Recognition Systems (pp.
145-180). New York: Springer Verlag.
Foo, K. C., Ngo, D. B. L., & Sagar, V. K. (1995, 18th October). Fuzzy Feature
Selection for Fingerprint Identification. Paper presented at the IEEE
International Carnahan Conference on Security Technology, Sanderstead,
U.K.
Ford, J. P., & Sticha, P. J. (1999). Introduction to Biometric Identification
Technology: Capabilities and Applications to the Food Stamp Program.
Arlington, Virginia: R. Lewis & Company.
Garris, M. D., Janet, S., McCabe, R., Tabassi, E., Watson, C. I., & Wilson, C. L.
(2004). User's Guide to NIST Fingerprint Image Software 2 (NFIS2).
Gaithesburg, MD: National Institute of Standards and Technology.
Grother, P. (2006, September). An Overview of ISO/IEC 19795-4. Paper
presented at the Biometric Consotium, Baltimore.
Grother, P., McCabe, M., Watson, C., Indovina, M., Salamon, W., Flanagan, P.,
et al. (2006). MINEX Performance and Interoperability of the INCITS 378
Fingerprint Template. Gaithersburg, Maryland: National Institute of
Standards and Technology.
Han, J., Kadowaki, T., Sato, K., & Shikida, M. (1999). Thermal Analysis of
Fingerprint Sensor Having a Microheater Array. Paper presented at the
International Symposium on Micromechatronics and Human Science,
Nagoya, Japan.
Han, Y., Nam, J., Park, N., & Kim, H. (2006). Resolution and Distortion
Compensation based on Sensor Evaluation for Interoperable Fingerprint
Recognition. Paper presented at the 2006 International Joint Conference
on Neural Networks, Vancouver, Canada.
Hawley, D., Roy, R., Yu, A. W., & Zhu, S. (1986). Frustrated Total Internal
Reflection: A demonstration and review. American Journal of Physics,
54(7), 602-607.
150
Haylock, S. E. (1979). Khan Bahadur Azizul Haque. Fingerprint World, 5(17), 2829.
Hong, L., Bolle, R., Jain, A. K., & Pankanti, S. (1997). An Identity Authentication
System Using Fingerprints. Proceedings of the IEEE, 85(9), 1365-1388.
Hutchings, P. J. (1997). Modern Forensics: Photography and Other Suspects.
Cardozo Studies in Law and Literature, 9(2), 229-243.
IBG. (2007). Biometrics Market and Industry Report. NY: IBG.
Igaki, S., Eguchi, S., Yamagishi, F., Ikeda, H., & Inagaki, T. (1992). Real-time
fingerprint sensor using a hologram. Applied Optics, 31(11), 1794-1802.
ISO. Biometric Data Interchange Formats -Part 2, Part4. Geneva: SO/IEC JTC1
SC37 WG3.
ISO. (2006). Text of 1st Working Draft 29794-1, Biometric Sample Quality
Standard - Part 1: Framework. Geneva: ISO/IEC JTC 1/SC37.
ISO/IEC JTC1 SC37 SD2 - Harmonized Biometric Vocabulary. (2006).).
Geneva: ISO/IEC.
Jain, A., Chen, Y., & Dass, S. (2005). Fingerprint Quality Indices for Predicting
Authentication Performance. Paper presented at the 5th International
Conf. on Audio- and Video-Based Biometric Person Authentication, Rye
Brook, NY.
Jain, A., Maltoni, D., Maio, D., & Wayman, J. (2005). Biometric Systems
Technology, Design and Performance Evaluation. London: Spring Verlag.
Jain, A., & Ross, A. (Eds.). (2004). Biometric Sensor Interoperability (Vol. 3067).
Berlin: Springer-Verlag.
Jain, L. C., Erol, A., & Halici, U. (1999). Introduction to Fingerprint Recognition. In
L. C. Jain, U. Halici, I. Hayashi, S. B. Lee & S. Tsutsui (Eds.), Intelligent
Biometric Techniques in Fingerprint and Face Recognition (pp. 1-34).
Boca Raton, Florida: CRC.
Jarosz, H., Fondeur, J. C., & Dupre, X. (2005). Large Scale Identification System
Design. In J. Wayman, A. Jain, D. Maltoni & D. Maio (Eds.), Biometric
Systems (pp. 263-288). London: Springer Verlag.
Jian, X., Lim, E., & Yau, W. (2002). Fingerprint quality and validity analysis.
Paper presented at the International Conference on Image Processing,
Rochester, N.Y.
Jiang, X. (2005). On Orientation and Anisotropy Estimation for Online Fingerprint
Authentication. IEEE Transactions on Signal Processing, 53(10), 40384049.
Khanna, R. (2004). Systems Engineering for Large-Scale Fingerprint Systems. In
N. Ratha & R. Bolle (Eds.), Automatic Fingerprint Recognition Systems
(pp. 283-304). New York: Springer.
Kim, H., Kang, H., Lee, B., Shin, D., & Kim, J. (2003). A Study on Performance
Evaluation of Fingerprint Sensors. Paper presented at the AVBPA 2003,
Guildford, UK.
151
Ko, T., & Krishnan, R. (2004). Monitoring and Reporting of Fingerprint Image
Quality and Match Accuracy for a Large User Application. Paper
presented at the Applied Imagery Pattern Recognition Workshop,
Washington, D.C.
Lim, T. W., & Moghavvemi, M. (2000). Capacitive Fingerprint Sesnsor Chip for
Automatic Matching. Paper presented at the TENCON 2000, Kuala
Lumpur, Malaysia.
M1. (2004). INCITS 378 - Finger Minutiae Format for Data Interchange.
Washington D.C.: ANSI/INCITS
Machida, K., Morimura, H., Okazaki, Y., & Shigematsu, S. (2004). Fingerprint
Sensor Technology- Sensor Structure/Chip, Sensing Scheme, and
System. Paper presented at the 7th International Conference on SolidState and Integrated Circuits Technology, Beijing, China.
Maio, D., & Maltoni, D. (1997). Direct Gray-Scale Minutaie Detection in
Fingerprints. IEEE Transactions on Pattern Analysis and Machine
Intelligence, 19(1), 27-40.
Mansfield, A., & Wayman, J. (2002). Best Practices. Middlesex: National Physics
Laboratory.
Marcialis, G., & Roli, F. (2004). Fingerprint Verification by Fusion of Optical and
Capacitive Sensors. Pattern Recognition Letters, 25(11), 1315-1322.
Mayers, J. C. (1788). Anatomical Copper-plates with Appropriate Explanations.
Mehtre, B. M. (1993). Fingerprint Image Analysis for Automatic Identification.
Machine Vision and Applications, 6, 124-139.
Montgomery, D. C. (1997). Design and Analysis of Experiments (4th ed.). New
York: John Wiley & Sons.
Muller, R. (2001). Fingerprint Verification with Microprocessor Security Tokens.
Munich University of Technology, Munich.
Nagdir, R., & Ross, A. (2006). A Calibration Model for Fingerprint Sensor
Interoperability. Paper presented at the SPIE Conference on Biometric
Technology for Human Identificaiton III, Orlando, USA.
O'Gorman, L. (1999). Biometrics Personal Identification in Networked Society.
Norwell: Kluwer Academic Publishers.
O'Gorman, L., & Xia, X. (2001). Innovations in fingerprint capture devices.
Pattern Recognition, 36(1), 361-369.
Poh, N., Kittler, J., & Bourlai, T. (2007, September). Improving Biometric Device
Interoperability by Likelihood Ratio-based Quality Dependent Score
Normalization. Paper presented at the BTAS 2007, Washington, D.C.
Polson, C. J. (1951). Finger Prints and Finger Printing: An Historical Study.
American Journal of Police Science, 41(5), 690-704.
Raafat, H., & Xiao, Q. (1991). Fingerprint Image Postprocessing: A Combined
Statistical and Structural Approach. Pattern Recognition, 24(10), 985-992.
Reed, B. (1981). Automated Fingerprint Identification: From Will West to
Minnesota Nine-Fingers and Beyond. Journal of Police Science and
Adminstration, 9, 318-319.
152
Secugen. SEIR Optic Technolgoy. Retrieved May 2, 2007, from
HTTP://WWW.SECUGEN.COM/DOWNLOAD/SGWP_SEIR.PDF
Sekran, U. (2003). Research Methods for Business. New York: John Wiley &
Sons, Inc.
Setlak, D. R. (2004). Advances in Fingerprint Sensors Using RF Imaging
Techniques. In N. Ratha & R. Bolle (Eds.), Automatic Fingerprint
Recognition Systems (pp. 27-53). New York: Springer-Verlag.
Shu-Long, J., Zhong-Kang, S., & Yan-Yan, W. (1991). The Non-Parametric
Detection with Neural Network. Paper presented at the International
Conference on Circuits and Systems, China.
Stock, R. M., & Swonger, C. W. (1969). Development and Evaluation of a Reader
of Fingerprint Minutiae. Ithica, N.Y. : Cornell Aeronautical Laboratory.
Tabassi, E., & Wilson, C. L. (2005). A novel approach to fingerprint image quality.
Paper presented at the International Conference on Image Processing,
Genoa, Italy.
Tabassi, E., Wilson, C. L., & Watson, C. I. (2004). Fingeprint Image Quality
(NISTIR 7151). Gaithersburg, MD: NIST.
Thieme, M. (2003). Effect of User Habituation in the Context of Biometric
Performance Testing. Washington D.C.: ANSI/INCITS.
Tolley, B. (2006). Notes for Using SAS & Homework Problems [Electronic
Version], 34. Retrieved 30th March, 2008 from
HTTP://WWW.UTMEM.EDU/PREVMED/PM/FILELIBRARY/SASNOTES2
006_2007.PDF.
TSA. (2006). Registered Traveler Program Business Model Version 2.1.
Washington D.C.: U.S. Department of Homeland Security.
Wayman, J. (1997). A Generalized Biometric Identification System Model. Paper
presented at the 31st Conference on Signals, Systems and Computers,
Pacific Grove, California.
Zhang, Y., Yang, J., & Wu, H. (2006). Coarse-to-fine image registration for
sweep fingerprint sensors. Optical Engineering, 45(6), 1-3.
153
Appendix A.
Figure A.1 Data Collection Protocol
154
Appendix B.
List of questions to be answered by each participant:
1. Age.
2. Gender.
3. Handedness.
4. Occupation.
5. Ethnicity.
Data Collected before first interaction with the fingerprint sensors:
1. Temperature in the lab.
2. Outdoor temperature.
Data Collected before interaction with a new fingerprint sensor:
1. Moisture content of skin on finger surface.
2. Oiliness of skin on finger surface.
3. Elasticity of skin on finger surface.
4. Temperature of skin on finger surface.
Data Collected during every interaction with the fingerprint sensors:
1. Peak pressure applied on the fingerprint sensor.
2. Failure to acquire.
155
Appendix C.
Figure C.1 Flowchart of fingerprint feature statistical analysis
156
Appendix D.
Figure D.1 Physical layout of data collection area
157
Appendix E.
Figure E.1 Normality plot of residuals of Aware Minutiae Count
Figure E.2 Residuals vs. Fitted values of Aware Minutiae Count
158
Figure E.3 Time series plot of observations
Figure E.4 Normality plot of residuals of NBIS Minutiae Count
159
Figure E.5 Residuals vs. Fitted values of NBIS Minutiae Count
Figure E.6 Time series plot of observations
160
Appendix F.
Figure F.1 Normality plot of residuals of Aware Quality Scores
161
Appendix G.
comparisons with S1 dataset
162
163
164
Appendix H.
165
166
167
Appendix I.
One hundred and ninety participants completed the survey and the fingerprint
collection activity. Fifty nine participants were females and 131 participants were
males. Two categories of labor distribution were created based on subjective
assessment of the amount of wear and tear an individual puts on their finger skin.
The two categories were manual laborer and office worker. Seventeen
participants placed themselves in the manual laborer group and 173 participants
placed themselves in the office worker group. Three age groups were created to
categorize the participants: less than 30 years, between 30 to 50 years, and 50
years and older. One hundred and fifty six participants were less than 30 years
old, 23 participants were between 30 and 50 years old, and 11 participants were
50 years and older. One hundred and sixty participants were right handed, 23
participants were left handed, and three participants were ambidextrous.
Ambidextrous participants were reassigned according to the dominant hand used
for writing. Six categories for ethnicity were used. One hundred and thirty three
participants identified themselves as Caucasian, eight as Black, one as American
Indian, 34 as Asian, 11 as Hispanic, and three participants as the Other category.
Table I.1 provides a summary of the results
168
Table I.1 Results of Survey
Total
190
Gender
Occupation
Age Groups
Male
Female
131
59
Manual Laborer
Office Worker
17
173
< 30
30-50 years
> 50 years
156
23
11
Right
Left
Ambidextrous
164
23
3
years
Handedness
Ethnicity
Caucasian
Black
Hispanic American
Asian
Other
34
3
Indian
133
8
11
1
169
Appendix J.
A correlation matrix was generated for examining the relation between quality
score of the fingerprint image, temperature, moisture content, oiliness and
elasticity of finger skin (Table J.1). The quality score was generated using Aware
quality software. The cells with the Pearson correlation coefficient are related to
the elements indicated in the corresponding row and column headers.
Table J.1 Correlation matrix of Quality Score, Temperature, Moisture Content,
Oiliness, Elasticity
Quality
Temperature
Score
Moisture
Oiliness
Content
Temperature
-0.006
Moisture Content
0.006
0.009
Oiliness
0.024
0.056
0.10
Elasticity
-0.20
0.051
0.09
-0.018
A scatter plot of quality score vs. temperature, quality score vs. moisture content,
quality score vs. oiliness and quality score vs. elasticity was generated. The data
points were grouped according to the type of acquisition technology used to
capture the fingerprint images. The scatter plot of quality score vs. moisture did
not show a specific relation but the biggest cluster of data points with the higher
quality scores was observed between 20 and 50 units on the moisture scale
(Figure J.2). The scatter plot of quality score vs. oiliness showed a negative
relation between quality score and oiliness units (Figure J.3). The scatter plot of
quality score vs. elasticity showed a positive relation between the quality score
and elasticity units (Figure J.4). These scatter plots indicated a difference in
relation between the finger skin characteristics and the quality scores for images
captured from different acquisition technologies. These scatter plots indicate that
170
a framework can be created to improve quality scores based on the finger skin
characteristics.
Figure J.1 Scatter plot of Aware Quality Score vs. Temperature
Figure J.2 Scatter plot of Aware Quality Score vs. Moisture
171
Figure J.3 Scatter plot of Aware Quality Score vs. Oiliness
Figure J.4 Scatter plot of Aware Quality Score vs. Elasticity
172
VITA
SHIMON K. MODI
EDUCATION
Doctorate of Philosophy, Technology, Expected August 2008.
Specialization in Computational Sciences & Engineering
Purdue University.
Dissertation: Fingerprint Sensor Interoperability: Analysis of Fingerprint
Sensor Interoperability on System Performance
Advisor: Dr. Stephen J. Elliott
Master of Science, Technology, May 2005.
Specialization in Information Security
Purdue University.
Thesis: Keystroke Dynamics Verification Using Spontaneous Password
Advisor: Dr. Stephen J. Elliott
Bachelor of Science, Computer Science, May 2003.
Dual Minors Management and Economics.
Purdue University
HONORS/AFFILIATIONS
Ross Fellowship for Doctoral Studies 2005-present
Semester Honors Spring ’02, Fall ’02, Spring ’03
STANDARDS INVOLVEMENT
•
•
•
•
ISO/IEC JTC1 SC-37 WG3 Data Interchange Formats
INCITS M1: Biometrics
Technical Editor for INCITS 1829-D: BioAPI Java Interfaces
INCITS M1.4 Ad-Hoc Group on Biometrics & E-Authentication
173
RESEARCH INTERESTS
•
•
•
Statistical Analysis and Testing of Biometrics Systems
Biometric sample quality assessment
Electronic authentication using biometrics, management of
federated identity systems, and integration of biometrics and
cryptography
INDUSTRY EXPERIENCE
US Biometrics – Biometrics Researcher (May 2007 - August 2007)
Designed and analyzed performance tests for biometric system
interoperability. The tests focused on integrating a different accquisition
subsytem into the existing security infrastructure. Also responsible for
requirements analysis for creating standards compliant physical access
and logical access biometrics system. Standards included FIPS-201,
BioAPI, and INCITS 378-2004.
RESEARCH EXEPERIENCE
Testing of Fingerprint Recognition Systems
Involved in diverse range of projects aimed at understanding impact of
image quality, impact of age, impact of environmental conditions, and
impact of finger preference on matching performance.
Biometrics and Cryptography
This is a collaborative research project with Center of Education and
Research in Information Assurance and Security (CERIAS) to incorporate
biometrics in cryptographic protocols like zero knowledge proofs.
Keystroke Dynamics Verification Using Spontaneous Password
Designed and tested multiple keystroke dynamics verification algorithms
for an authentication mechanism that used a spontaneously generated
password for verification instead of a pre-known, static password.
Biometrics in Federated Architecture Systems
Joint research with Center of Education and Research in Information
Assurance and Security (CERIAS) examining methodologies to
incorporate biometrics into federated architecture system, and other
identity management system architectures.
174
Equine Iris Recognition – Iristrac™
This research examined the feasibility of using equine eye features for
recognition. Extracting equine eye features, development of pattern
matching algorithms, and feasibility testing were the main research goals.
Security Analysis of Different Biometrics Storage/Matching
Architectures– NIST
Created a technical report for National Institute of Standards in
Technology (NIST) which examined architectures and interoperability of
current biometric systems. This research also analyzed the current
methodologies for revocation of a compromised biometric identifier.
PUBLICATIONS
Chapter Co-Author:
• Elliott, S. J., Kukula, E., & Modi, S. (2007). Issues Involving the
Human Biometric Sensor Interface. In S. Yanushkevich, P. Wang &
S. Srihari (Eds.), Image Pattern Recognition Synthesis and
Analysis in Biometrics (Vol. 67, pp. 400): World Scientific
Publishers.
Co-Author for the following locally published books:
• Introduction to Biometric Technology.
• Securing the Manufacturing Environment Using Biometric
Technologies.
Conference Proceedings
•
•
•
•
•
Modi, S. K., Elliott, S. J., Kim, H., & Kukula, E. P. (2008). Statistical
Analysis Framework for Biometric System Interoperability Testing.
Paper presented at the ICITA 2008, Cairns, Australia.
Frick, M. D., Modi, S. K., Elliott, S. J., & Kukula, E. P. (2008).
Impact of Gender on Fingerprint Recognition Systems. Paper
presented at the ICITA 2008, Cairns, Australia.
Bhargav-Spantzel, A., Modi, S., Bertino, E., Elliott, S. J., Young, M.,
& Squicciarini, A. C. (2007). Privacy Preserving Multi-Factor
Authentication with Biometrics. Journal of Computer Science
Modi, S. K., Elliott, S. J., & Kim, H. (2007 ). Performance Analysis
for Multi Sensor Fingerprint Recognition System. Paper presented
at the ICISS 2007, New Delhi, India
Modi, S., Elliott, S. J., Kim, H., & Whetstone, J. (2007). Impact of
Age Groups on Fingerprint Recognition Performance. Paper
presented at the AutoID 2007, IEEE Workshop on Automatic
Identification Advanced Technologies, Alghero, Italy.
175
•
•
•
•
Modi, S. K., & Elliott, S. J. (2006). Impact of Image Quality on
Performance: Comparison of Young and Elderly Fingerprints.
Paper presented at the 6th International Conference on Recent
Advances in Soft Computing (RASC), Canterbury, UK.
Modi, S. K., & Elliott, S. J. (2006). Graduate Course Development
focusing on Security Issues for Professionals Working in the
Manufacturing Industry. Paper presented at the World Congress on
Computer Science, Engineering and Technology Education,
Santos, Brazil.
Modi, S. K., & Elliott, S. J. (2006). Keystroke Dynamics Verification
Using Spontaneously Generated Password. Paper presented at the
40th IEEE International Carnahan Conferences Security
Technology, Lexington, Kentucky.
Modi, S. K., & Elliott, S. J. (2005). Securing the Manufacturing
Environment using Biometrics. Paper presented at the 39th Annual
International Carnahan Conference on Security Technology
(ICCST), Las Palmas de G. C., Spain

Design Guidelines for Deploying Closed Loop Systems

Transcription

Similar documents

Seamless integration

Operation Kidsafe Safety Week Flyer Vaden Nissan 2016

CSE 40537 / 60537: Biometrics

R3310 POS Series

BioDeadbolt

64Bit Processor --- KINGZONE N3 PLUS Is Coming

Why I Hacked TouchID and Still Think It`s Awesome

Session 1 (Pannuto) - USAID Learning Lab